Measurement system, measurement system anomaly determination method, and measurement system anomaly determination program

Abstract

This measurement system is provided with: a measurement instrument 20 which has at least a first sensor 21 and a second sensor 22; a predicted value estimation unit 35 which uses a machine learning model that has learned the relation between measured values of the first sensor 21 and measured values of the second sensor 22 to estimate a predicted value of the second sensor 22 from a measured value of the first sensor 21; and an anomaly determination unit 36 which compares the predicted value of the second sensor 22 and the measured value of the second sensor 22 and determines whether an anomaly is present in the machine learning model or in the measurement instrument 20.

Description

Measurement system, method for determining abnormality in a measurement system, and program for determining abnormality in a measurement system

[0001] The present invention relates to a measurement system for measuring the properties of a sample such as water quality, a method for determining an abnormality in the measurement system, and a program for determining an abnormality in the measurement system.

[0002] For example, in a liquid analysis device such as that described in Patent Document 1, which is equipped with a sensor for measuring water quality, conventionally, if an obvious abnormality occurs, such as the sensor output value going off the scale, the user can determine that maintenance is required for the sensor.

[0003] However, even if there is no obvious abnormality in the sensor, the sensor output value may deviate from the value that should be output due to factors such as sensor dirt or a malfunction of a component that affects the sensor output value.

[0004] In such cases, it is difficult for the user to determine whether the sensor output value has changed due to dirt or a malfunction of the sensor or other components.

[0005] JP 2015-25794 A

[0006] In recent years, it has also been considered to use machine learning to automatically determine whether or not a measurement system needs maintenance. In this machine learning, supervised learning is performed using a combination of past sensor output values ​​and measurements obtained from those output values ​​as training data (training data sets). A possible method is to compare the predicted values ​​obtained by this machine learning with the actual measured values ​​to determine the need for maintenance.

[0007] However, if unknown data that deviates significantly from the training dataset is input due to changes in sample trends, the predicted values ​​estimated by the machine learning model may differ significantly from the actual measured values. This may result in the system determining that the measurement system requires maintenance even when it does not.

[0008] Therefore, the present invention has been made to solve the above-mentioned problems, and its main objective is to determine the need for maintenance of measuring equipment or the need for re-learning of machine learning models in a measurement system that utilizes machine learning.

[0009] That is, the measurement system according to the present invention is characterized by comprising a measuring device having at least a first sensor and a second sensor, a predicted value estimation unit that estimates a predicted value of the second sensor from the actual measured value of the first sensor using a machine learning model that has learned the relationship between the actual measured value of the first sensor and the actual measured value of the second sensor, and an abnormality determination unit that compares the predicted value of the second sensor with the actual measured value of the second sensor to determine whether an abnormality exists in the machine learning model or the measuring device.

[0010] According to the measurement system configured in this manner, the predicted value of the second sensor is estimated from the actual measured value of the first sensor using a machine learning model, and the estimated predicted value of the second sensor is compared with the actual measured value of the second sensor to determine whether an abnormality exists in the machine learning model or the measuring device. This makes it possible to accurately determine the need for maintenance of the measuring device. Furthermore, if the machine learning model determines that an abnormality exists, it is possible to detect that the tendency of the sample has changed since the machine learning model was created, and it is also possible to re-train the machine learning model.

[0011] As a specific implementation mode of the abnormality determination, it is desirable that the abnormality determination unit determines whether the difference or ratio between the predicted value of the second sensor and the actual measured value of the second sensor is within a predetermined tolerance range, and if it is outside the tolerance range, determines whether there is an abnormality in the machine learning model or the measuring device.

[0012] Specifically, it is desirable that the anomaly determination unit uses a non-parametric density estimation method to determine whether an anomaly exists in the machine learning model or the measuring device.

[0013] In a specific embodiment of anomaly determination using nonparametric density estimation, the anomaly determination unit preferably determines that the machine learning model is abnormal when the data density of the learning dataset from which the machine learning model was created is determined to be abnormal in a region where the data density is equal to or greater than a predetermined threshold, and determines that the measuring device or the machine learning model is abnormal when the data density is determined to be abnormal in a region where the data density is less than the threshold. Note that when the measuring device or the machine learning model is determined to be abnormal in a region where the data density is less than the threshold, if an anomaly is not confirmed by analyzing internal data of the measuring device, it can determine that the machine learning model is likely to be abnormal.

[0014] Furthermore, it is desirable that the measurement system according to the present invention further includes a machine learning unit that creates the machine learning model using a training dataset consisting of the actual measurement values ​​of the first sensor and the actual measurement values ​​of the second sensor. By including the machine learning unit in this measurement system, it is possible to create a machine learning model using a training dataset obtained from a sample measured by a user.

[0015] It is desirable that the machine learning unit retrains the machine learning model when the anomaly determination unit determines that the machine learning model is abnormal. By retraining the machine learning model in this manner, the reliability of the machine learning model can be improved. Then, by estimating the predicted value of the second sensor using the retrained machine learning model, the need for maintenance of the measuring device can be determined with even greater accuracy.

[0016] In order to prompt the user to perform maintenance, it is desirable that when the abnormality determination unit determines that the measuring device is abnormal, a maintenance signal is output to prompt the user to perform maintenance on the measuring device.

[0017] Each of the first sensor and the second sensor may measure any of the conductivity, resistivity, oxidation-reduction potential (ORP), chemical oxygen demand (COD), turbidity, color, temperature, pressure, oil film, or concentration of a predetermined component contained in the sample.

[0018] Here, users periodically perform sensor maintenance, which is a series of operations including cleaning the sensor, calibrating the sensor, replacing the sensor, relearning the machine learning model, and changing the maintenance interval. Conventionally, during sensor maintenance, the deterioration trend, which is the change in sensor deterioration over time, is identified, and a maintenance interval reflecting the deterioration trend is calculated for each sensor. On the other hand, a method of calculating a maintenance interval reflecting the deterioration trend for each sensor requires identifying the deterioration trend for each of multiple sensors. In the same measurement system, the deterioration trends of multiple sensors belonging to the same measurement system may be similar. The term "same measurement system" as used herein refers, for example, to a flow path within the same tank or water treatment facility that has no inflow from other sources from upstream to downstream. In this case, once a user identifies the deterioration trend of a certain sensor, they may estimate the deterioration trends of other sensors based on the deterioration trend of the sensor.

[0019] In order to address the above, the measurement system further includes a degradation determination unit that determines whether the difference between the actual state value and the predicted state value of the first sensor is within an acceptable range based on reference degradation data that indicates the relationship between a predicted state value of the first sensor and time and actual degradation data that indicates the relationship between an actual state value of the first sensor and time, and when the degradation determination unit determines that the difference is outside the acceptable range, it outputs an out-of-tolerance degradation signal that indicates that the state of the first sensor has deteriorated outside the acceptable range.

[0020] In this way, the user recognizes the out-of-tolerance degradation signal output by the measurement system, which indicates out-of-tolerance degradation of the first sensor, and can infer that out-of-tolerance degradation is also occurring in other sensors in the measurement system, allowing the user to check the degradation status of other sensors in the same measurement system even if the other sensors are not within the pre-calculated maintenance intervals.

[0021] The measurement system may further include a measured deterioration data calculation unit that calculates first measured deterioration data indicating the relationship between the actual condition value of the first sensor and time and second measured deterioration data indicating the relationship between the actual condition value of the second sensor and time, and a maintenance cycle calculation unit that calculates the maintenance cycle of the first sensor and the maintenance cycle of the second sensor based on the first measured deterioration data and the second measured deterioration data.

[0022] In this way, the measurement system calculates the maintenance cycles for the first sensor and the second sensor based on the first measured deterioration data and the second measured deterioration data, so that the maintenance cycles for the first sensor and the second sensor can be changed to suit the actual condition measurements.

[0023] A specific embodiment of the present invention is one in which the measurement system is used for water quality analysis.

[0024] In addition, the method for determining an abnormality in a measurement system according to the present invention is a method for determining an abnormality in a measurement system that measures a sample using at least a first sensor and a second sensor, and is characterized in that a predicted value of the second sensor is estimated from the actual measured value of the first sensor using a machine learning model that has learned the relationship between the actual measured value of the first sensor and the actual measured value of the second sensor, and the predicted value of the second sensor is compared with the actual measured value of the second sensor to determine whether there is an abnormality in the machine learning model or the measurement equipment.

[0025] Furthermore, the abnormality determination program for a measurement system according to the present invention is an abnormality determination program for a measurement system having at least a first sensor and a second sensor, and is characterized in that it causes a computer to perform the functions of a predicted value estimation unit that estimates a predicted value of the second sensor from the actual measured value of the first sensor using a machine learning model that has learned the relationship between the actual measured value of the first sensor and the actual measured value of the second sensor, and an abnormality determination unit that compares the predicted value of the second sensor with the actual measured value of the second sensor and determines whether there is an abnormality in the machine learning model or the measuring device.

[0026] According to the present invention, in a measurement system that utilizes machine learning, it is possible to determine the need for maintenance of a measurement device or the need for re-learning of a machine learning model.

[0027] 1 is a schematic diagram showing the configuration of a measurement device according to a first embodiment of the present invention; FIG. 2 is a diagram showing the physical configuration of an information processing device according to the first embodiment; FIG. 3 is a diagram showing functional blocks of an information processing device according to the first embodiment; FIG. 4 is a flowchart of an abnormality determination method according to the first embodiment; FIG. 5 is a diagram showing (a) a graph showing a learning dataset of sensors A and B, and (b) kernel density according to the first embodiment; FIG. 6 is an analysis graph of sensors A and B according to the first embodiment; (a) an analysis graph of sensors A and B, (b) an analysis graph of sensors A and C, and (c) an analysis graph of sensors B and C; FIG. 7 is a diagram showing functional blocks of an information processing device according to a second embodiment; FIG. 8 is a flowchart of a degradation determination method according to the second embodiment; FIG. 9 is a graph of predicted state values ​​and actual state measured values ​​relative to the sensitivities of the first and second sensors according to the second embodiment;

[0028] First Embodiment A first embodiment, which is an embodiment of a measurement system according to the present invention, will be described below with reference to the drawings.

[0029] <Basic Configuration of Measurement System> The measurement system 100 according to this embodiment measures one or more measurement items of a liquid sample such as clean water or sewage in, for example, a water treatment facility.

[0030] Here, the measurement items include, for example, hydrogen ions (H ＋ ), ammonium nitrogen (NH ４ -N), residual chlorine, dissolved oxygen (DO), salinity, fluoride ions, nitrate ions (NO ３ － ), hydrogen fluoride (HF), potassium hydroxide (KOH), tetramethylammonium hydroxide (TMAH), dissolved ozone, phosphoric acid (H ３ P.O. ４ ), citric acid, ammonia (NH ３ ), nitrogen (N ２), phosphorus (P), or silica (SiO ２ Examples of indicators of water quality include the concentration of a specific component such as the concentration of a specific component of the sample, the conductivity, resistivity, oxidation-reduction potential (ORP), chemical oxygen demand (COD), turbidity, color, temperature, pressure, or oil film of the sample.

[0031] Specifically, as shown in FIG. 1, the measurement system 100 includes a measurement device 20 having at least a first sensor and a second sensor, and an information processing device 30 that processes information output from the measurement device 20.

[0032] <Configuration of Measuring Device 20> The measuring device 20 measures one or more measurement items and has multiple sensors 21-23 corresponding to each measurement item. While FIG. 1 shows the measuring device 20 having three different sensors 21-23, it may have two different sensors, or four or more different sensors. Furthermore, the measuring device 20 may have multiple identical sensors installed at different measurement locations. Hereinafter, the three sensors 21-23 will also be referred to as the first sensor 21, the second sensor 22, and the third sensor 23.

[0033] The sensors 21-23 of the measuring device 20 can be changed as appropriate depending on the purpose of measurement. Examples of the sensors 21-23 of the measuring device 20 include an oxidation-reduction potential sensor, a pH sensor, a DO sensor, a sludge turbidity sensor (MLSS measuring device), an ammonia concentration sensor, and a temperature sensor. The multiple sensors 21-23 of the measuring device 20 may be a single measuring device unitized by being held in a single casing, or may be separate measuring devices installed independently of each other. When the sensors are separate measuring devices, they may be located in different locations, such as different aeration tanks of a water treatment facility or different locations on a sample pipe.

[0034] The output values ​​output from the sensors 21 to 23 of the measuring device 20 configured in this manner are processed by the information processing device 30. The output values ​​of the sensors 21 to 23 can be transmitted to the information processing device 30 via wired or wireless communication.

[0035] <Configuration of information processing device 30> As shown in FIG. 2, the information processing device 30 is a computer having a CPU 30a, a memory 30b, an input / output interface 30c, an AD converter 30d, an input means 30e such as a keyboard, a display means 30f such as a display, and a communication means 30g for communicating via a network.

[0036] This information processing device 30 performs the functions of a calculation unit 31, a display control unit 32, and a memory unit 33, as shown in Figure 3, by the CPU 30a and peripheral devices working together based on a program stored in a specified area of ​​the memory.

[0037] The calculation unit 31 calculates the measurement value of each measurement item based on the output value of each sensor 21 to 23 of the measuring device 20. The measurement values ​​of each sensor 21 to 23 calculated by the calculation unit 31 are sent to the display control unit 32 and the storage unit 33.

[0038] The display control unit 32 displays on the display 30f the output values ​​or measurement values ​​of each of the sensors 21 to 23 of the measuring device 20. Specifically, the display control unit 32 can, for example, display on the display 30f a graph with one axis (horizontal axis) representing time and the other axis (vertical axis) representing the output value or measurement value, and can display on this graph the change over time in the output values ​​or measurement values ​​from each of the sensors 21 to 23. The display control unit 32 can also display a list of measurement values, such as instantaneous values, of each of the sensors 21 to 23 of the measuring device 20.

[0039] The storage unit 33 stores output values ​​or measurement values ​​of the sensors 21 to 23 of the measuring device 20. The various data stored in the storage unit 33 can be displayed on the display 30f by the display control unit 32.

[0040] <Abnormality Determination Function Using Machine Learning> Furthermore, as shown in Fig. 3, the information processing device 30 functions as a machine learning unit 34, a predicted value estimation unit 35, and an abnormality determination unit 36. Below, the functions of the machine learning unit 34, the predicted value estimation unit 35, and the abnormality determination unit 36 ​​will be described with reference to the flowchart of the abnormality determination method shown in Fig. 4. Furthermore, for convenience of explanation, of the three sensors 21 to 23, two sensors, the first sensor 21 and the second sensor 22, will be described, but the same applies to the two sensors, the second sensor 22 and the third sensor 23, and the two sensors, the first sensor 21 and the third sensor 23.

[0041] The machine learning unit 34 uses a learning dataset consisting of past actual measurement values ​​of the first sensor 21 and past actual measurement values ​​of the second sensor 22 to create a machine learning model that learns the relationship between the actual measurement values ​​of the first sensor 21 and the actual measurement values ​​of the second sensor 22.

[0042] The learning algorithm of the machine learning unit 34 may be a support vector machine, a self-organizing map, an artificial neural network, a decision tree, a random forest, k-means, k-nearest neighbors, a genetic algorithm, a Bayesian network, or a deep learning method.

[0043] Here, the past actual measurement values ​​of the first sensor 21 are past output values ​​or measurement values ​​of the first sensor 21, and the past actual measurement values ​​of the second sensor 22 are past output values ​​or measurement values ​​of the second sensor 22. Note that the learning dataset may be a combination of the past output values ​​of the first sensor 21 and the past output values ​​of the second sensor 22, a combination of the past measurement values ​​of the first sensor 21 and the past measurement values ​​of the second sensor 22, or a combination of the past output values ​​of one sensor and the past measurement values ​​of the other sensor.

[0044] The predicted value estimation unit 35 estimates the predicted value of the second sensor 22 from the actual measured value of the first sensor 21 using the machine learning model created by the machine learning unit 34. The predicted value of the second sensor 22 is an output value or measurement value of the second sensor 22 that is predicted to be obtained at the same time as the actual measured value of the first sensor 21 is measured.

[0045] The abnormality determination unit 36 ​​compares the predicted value of the second sensor 22 estimated by the predicted value estimation unit 35 with the actual measured value obtained by the second sensor 22 actually measuring the sample, and determines whether there is an abnormality in the machine learning model or the measuring equipment 20.

[0046] Specifically, the abnormality determination unit 36 ​​determines whether the difference or ratio between the predicted value of the second sensor 22 and the actual measured value of the second sensor 22 is within a predetermined tolerance, and if it is outside the tolerance, determines whether the abnormality is in the machine learning model or the measuring device 20 ("Alert Generation" in FIG. 4). The tolerance can be set in advance as, for example, a range of ±10% of the predicted value.

[0047] Here, the anomaly determination unit 36 ​​uses a nonparametric density estimation method to determine whether an anomaly exists in the machine learning model or the measuring device 20. As the nonparametric density estimation method, for example, a kernel density estimation method may be used.

[0048] More specifically, as shown in FIG. 5(b), if the abnormality determination unit 36 ​​determines that an abnormality exists in the area (inside the kernel density) where the data density of the learning dataset from which the machine learning model was created is equal to or greater than a predetermined threshold value (e.g., 0.3), the abnormality determination unit 36 ​​determines that an abnormality exists in the machine learning model ("inside the kernel density" in FIG. 4).

[0049] If the anomaly determination unit 36 ​​determines that the machine learning model has an abnormality, it outputs a re-learning signal to prompt the user to re-learn the machine learning model. Here, the re-learning signal is displayed, for example, on a display. When the user inputs a re-learning command, the machine learning unit 34 re-learns the machine learning model using the updated learning data set (the newly accumulated learning data set) and sets a threshold value for determining an abnormality in the anomaly determination unit 36 ​​("Re-learning model ⇒ reset threshold value" in FIG. 4). Note that each threshold value of the anomaly determination unit 36 ​​may be set arbitrarily by the user.

[0050] On the other hand, as shown in FIG. 5( b), if the abnormality determination unit 36 ​​determines that an abnormality exists in a region where the data density is less than the threshold (outside the kernel density) ("outside the kernel density" in FIG. 4), the abnormality determination unit 36 ​​determines that an abnormality exists in the machine learning model or the measuring device 20. Then, the abnormality determination unit 36 ​​checks whether an abnormality exists in the first sensor 21 or the second sensor 22 ("Check each sensor" in FIG. 4). Note that if no abnormality is confirmed in the first sensor 21 or the second sensor 22, the abnormality determination unit 36 ​​determines that an abnormality exists in the machine learning model, and outputs a re-learning signal to prompt the user to re-learn the machine learning model, as described above.

[0051] If the abnormality determination unit 36 ​​determines that the first sensor 21 or the second sensor 22 is abnormal, it outputs a maintenance signal to prompt the user to perform maintenance on the measuring device 20 ("Maintenance Recommendation" in FIG. 4). For example, the abnormality determination unit 36 ​​integrates the data density and the difference between the predicted value and the actual measurement value, compares the integrated value (data density x (predicted value - actual measurement value)) with a predetermined maintenance threshold, and outputs the maintenance signal. Here, the maintenance signal is displayed on a display, for example, and may be a signal indicating that the measuring device 20 is abnormal or a signal indicating that maintenance is required. This prompts the user to perform maintenance on the measuring device 20 that has been determined to be abnormal ("Device Maintenance" in FIG. 4).

[0052] Here, if an abnormality is detected by the abnormality determination unit 36, the following processing can also be performed. For example, as shown in Fig. 6, time-series data consisting of output values ​​or actual measurement values ​​of two sensors (sensor A and sensor B) selected from the plurality of sensors 21 to 23 can be plotted on a graph (hereinafter referred to as an analysis graph) with two sensor axes (sensor A on the horizontal axis and sensor B on the vertical axis), and data (learning data) from the two sensors (sensor A and sensor B) included in the learning dataset can also be plotted on the analysis graph.

[0053] In this way, by plotting the actual output values ​​or measurement values ​​(measured data) of two sensors (sensor A and sensor B) and the data (learning data) of the two sensors (sensor A and sensor B) on an analysis graph, it is possible to infer the cause of an incorrect prediction result. For example, if the measured data is plotted in an area with little learning data, it is possible to infer that the sample trend has changed. Furthermore, if only the measured data of a specific sensor has changed, it is possible to infer that an abnormality in that specific sensor exists.

[0054] In a configuration having three or more sensors, it is also possible to display analysis graphs of all patterns selected from the sensors, as shown in FIG. 7. For example, in a configuration having sensors A, B, and C, an analysis graph for sensors A and B, an analysis graph for sensors A and C, and an analysis graph for sensors B and C are displayed. Here, for example, by placing the cursor over one of the measured data in one analysis graph and selecting it, the other two analysis graphs can display measured data from the same time as the selected measured data. In this way, by comprehensively examining the measured data from three or more sensors, it is possible to analyze phenomena that cannot be analyzed using the measured data from two sensors.

[0055] Furthermore, time series data can be displayed on an analysis graph with two sensor axes by day of the week, month, or season, making it possible to analyze the influence of the day of the week (such as the factory not operating on Saturdays and Sundays), the influence of the month (such as busy or slow seasons), and the influence of the season (such as busy or slow seasons).

[0056] Effect of First Embodiment According to the measurement system 100 configured in this manner, a machine learning model is used to estimate a predicted value of the second sensor 22 from an actual measurement value of the first sensor 21, and the estimated predicted value of the second sensor 22 is compared with the actual measurement value of the second sensor 22 to determine whether an abnormality exists in the machine learning model or the measuring device 20. This makes it possible to accurately determine the need for maintenance of the measuring device 20. Furthermore, if the machine learning model determines that an abnormality exists, it is possible to detect a change in the tendency of the sample and to retrain the machine learning model.

[0057] Second Embodiment Next, a second embodiment of the present invention will be described with reference to the drawings. The same reference numerals will be used to denote the functional units described in the first embodiment. That is, the measuring device 20 and the information processing device 30 have the following additional functions in addition to the functions described in the first embodiment.

[0058] <Device Configuration> The measurement system 100 in the second embodiment of the present invention is preferably intended for the same measurement system S, such as a flow path from upstream to downstream within the same tank or water treatment facility, with no inflow from other sources, but is not limited to this.

[0059] The sensors 21 to 23 in the measuring device 20 measure measurement items in the same measurement system S. Of these, the first sensor 21 is preferably a sensor that is not affected by the optical system, such as a pH meter, but is not limited to this. The output values ​​output from the sensors 21 to 23 are processed by the information processing device 30.

[0060] The information processing device 30 functions as a display control unit 32 and a storage unit 33 as shown in FIG. 8, by the CPU 30a and peripheral devices working together based on a program stored in a predetermined area of ​​the memory 30b.

[0061] The display control unit 32 displays the output values, status values, or signals of each of the sensors 21 to 23 on a display. Note that the status values ​​referred to here are parameters that indicate the status of the sensor, such as the sensitivity of the sensor. Specifically, as shown in FIG. 10, the display control unit 32 can display, for example, a graph on the display 30f, with one axis (horizontal axis) representing time and the other axis (vertical axis) representing sensitivity. The display control unit 32 can also display the signals emitted from each of the sensors 21 to 23 on the display 30f.

[0062] The memory unit 33 stores the output values, status values, or maintenance cycles of each sensor 21 to 23, and also stores reference deterioration data D that indicates the relationship between the status predicted value, which is the predicted value of the status value that predicts the status of each sensor 21 to 23, and time.

[0063] <Deterioration Determination Function> In this embodiment, if the status value of a certain sensor falls below a limit value (e.g., 50%), the sensor will fail to measure properly and will need to be replaced. Therefore, a standard maintenance cycle, which is a maintenance cycle corresponding to the standard deterioration data D, is defined for each of the sensors 21 to 23, and the user performs maintenance on each sensor at each standard maintenance cycle and checks the deterioration state of each sensor. If it is predicted that the status value of a sensor will fall below the limit value before the standard maintenance cycle is reached, the user will replace the sensor before maintenance.

[0064] 8, the information processing device 30 further has functions as a state measurement value calculation unit 301, a deterioration determination unit 302, a measured deterioration data calculation unit 303, and a maintenance cycle calculation unit 304. Below, the functions of the state measurement value calculation unit 301, the deterioration determination unit 302, the measured deterioration data calculation unit 303, and the maintenance cycle calculation unit 304 will be described with reference to the flowchart shown in FIG. 9. For ease of explanation, of the sensors 21 to 23, only two sensors, the first sensor 21 and the second sensor 22, will be described, but the same applies to the two sensors, the second sensor 22 and the third sensor 23, and the two sensors, the first sensor 21 and the third sensor 23.

[0065] As shown in FIG. 9 , the user starts maintenance of the first sensor 21 at the reference maintenance cycle T1. First, the user cleans and calibrates the first sensor 21. When the state value of the calibrated first sensor 21 is output, the output state value is sent to the state measurement value calculation unit 301. The second sensor 22 and the third sensor 23 are cleaned and calibrated at the time of their respective maintenance. However, the embodiment is not limited to this, and the cleaning and calibration of the second sensor 22 and the third sensor 23 may be performed at the time of maintenance of the first sensor 21.

[0066] Next, the state measurement value calculation unit 301 calculates the state measurement value Y1a obtained by measuring the state of the first sensor 21 based on the output value output from the first sensor 21. Specifically, when the state value of the first sensor 21 at the time of calibration is output, the state measurement value calculation unit 301 calculates the state measurement value Y1a corresponding to the output state value (see FIG. 10 ). At this time, if the state measurement value Y1a is equal to or less than the limit value, the user replaces the first sensor 21 after calibration, and the maintenance is completed. On the other hand, if the state value of the first sensor 21 is greater than the limit value, the user does not replace the first sensor 21, and the obtained state measurement value Y1a is sent to the memory unit 33 for storage and also sent to the deterioration determination unit 302.

[0067] The deterioration determination unit 302 compares the predicted state value Y1 and the actual state measurement value Y1a of the first sensor 21 to determine the state of the first sensor 21. Specifically, as shown in Fig. 10, the deterioration determination unit 302 subtracts the actual state measurement value Y1a calculated by the actual state measurement value calculation unit 301 from the predicted state value Y1 stored in the storage unit 33, and determines whether the difference (Y1-Y1a) is within an allowable range. The allowable range can be set in advance to, for example, a range of ±10% of the predicted state value Y1.

[0068] If the difference (Y1-Y1a) is within the allowable range, the deterioration determination unit 302 determines that the deterioration of the first sensor 21 is within the allowable range. In this case, the user does not change the maintenance cycle of the first sensor 21, and also assumes that the deterioration of the other sensors is within the allowable range. The user then calibrates or cleans each of the sensors 21 to 23, and completes the maintenance.

[0069] On the other hand, if the difference (Y1-Y1a) is outside the allowable range, the degradation determination unit 302 determines that the degradation is outside the allowable range and outputs an out-of-allowable degradation signal indicating that the first sensor 21 is degraded outside the allowable range. Here, the out-of-allowable degradation signal is displayed, for example, on the display 30f of the display control unit 32. The out-of-allowable degradation signal is also output to the memory unit 33 and the measured degradation data calculation unit 303. The user is then prompted to check the status value of the second sensor 22 in the same measurement system S by recognizing the out-of-allowable degradation signal displayed on the display 30f. The status value of the second sensor 22 is then output to the state measurement value calculation unit 301, which then calculates the state measurement value Y2a of the second sensor 22. The calculated state measurement value Y2a is sent to the memory unit 33 and the measured degradation data calculation unit 303.

[0070] 10, an example of a case where degradation is determined to be outside the allowable range is when the actual condition value Y1a is lower than the predicted condition value Y1, but this is not limited to this. That is, even if the actual condition value Y1a is higher than the predicted condition value Y1, the degradation determination unit 302 determines that degradation is outside the allowable range. In this case, the actual condition value Y1a at the standard maintenance cycle T1 is better than the predicted condition value Y1, that is, the actual rate of degradation is slower than the standard degradation data D1.

[0071] The measured deterioration data calculation unit 303 calculates first measured deterioration data M1 indicating the relationship between the measured state value of the first sensor 21 and time, and second measured deterioration data M2 indicating the relationship between the measured state value of the second sensor 22 and time. Specifically, as shown in FIGS. 9 and 10 , upon receiving the out-of-tolerance deterioration signal from the deterioration determination unit 302, the measured deterioration data calculation unit 303 calculates the first measured deterioration data M1 based on the measured state value Y1a corresponding to the first sensor 21 and the past measured state values ​​of the first sensor 21 stored in the memory unit 33. The calculated first measured deterioration data M1 is then sent to the memory unit 33 and the maintenance cycle calculation unit 304. The calculation of the second measured deterioration data M2 for the second sensor 22 is similar to that for the first sensor 21.

[0072] The maintenance cycle calculation unit 304 calculates the maintenance cycles of the first sensor 21 and the second sensor 22 based on the measured deterioration data of the first sensor 21 and the second sensor 22. Specifically, as shown in Fig. 10, the maintenance cycle calculation unit 304 calculates the measured maintenance cycle T1a that will reach the same value as the predicted condition value Y1 based on the first measured deterioration data M1. The calculation of the measured maintenance cycle T2a for the second sensor 22 is similar to that for the first sensor 21.

[0073] The obtained measured maintenance periods T1a and T2a are sent to the display control unit 32 and the storage unit 33. The user then changes the measured maintenance periods T1a and T2a displayed on the display control unit 32 to new maintenance periods, and the maintenance is completed.

[0074] 11 , by comparing the reference maintenance periods T1 and T2 with the obtained actual maintenance periods T1a and T2a, and based on the change rates (T1a / T1, T2a / T2) between them, the user can estimate that the deterioration trends representing the changes over time in the status values ​​of the same measurement system S are similar for other sensors in the same measurement system S. In this case, the user can estimate, for example, the actual maintenance period T3a of the third sensor 23 in the same measurement system S from the change rate (T1a / T1 or T2a / T2) of the maintenance period of the first sensor 21 or the second sensor 22.

[0075] Furthermore, based on the rate of change (T1a / T1) of the maintenance period of the first sensor 21, the user can estimate the actual maintenance periods T2a and T3a of the second sensor 22 and the third sensor 23. Specifically, the user can estimate the actual maintenance periods T2a and T3a of the second sensor 22 and the third sensor 23 by multiplying the rate of change (T1a / T1) of the maintenance period of the first sensor 21 by the reference maintenance periods T2 and T3 of the second sensor 22 and the third sensor 23, respectively.

[0076] On the other hand, to improve the accuracy of the maintenance cycle estimation, the user can estimate the actual maintenance cycle T3a of the third sensor 23 based on both the rate of change (T1a / T1) of the maintenance cycle of the first sensor 21 and the rate of change (T2a / T2) of the maintenance cycle of the second sensor 21. Specifically, the user can estimate the actual maintenance cycle T3a of the third sensor 23 by calculating the average of the two rate of change (T1a / T1 or T2a / T2) and multiplying this by the reference maintenance cycle T3 for the third sensor.

[0077] In the above embodiment, the deterioration determination unit 302 calculates the difference (Y1-Y1a) to determine the deterioration state of the first sensor 21, but the user may calculate the difference (Y1-Y1a) to determine the deterioration state of the first sensor 21. In this case, when the user determines that the deterioration is outside the allowable range, the user outputs an outside-allowable-range deterioration signal to the measured deterioration data calculation unit 303.

[0078] Effect of the Second Embodiment In the measurement system 100 configured as described above, the deterioration determination unit 302 calculates the difference between the actual state value Y1a and the predicted state value Y1 of the first sensor 21. If the difference is outside the allowable range, an out-of-allowable deterioration signal is output, prompting the user to check the deterioration state of other sensors in the same measurement system S. Furthermore, the maintenance cycle calculation unit 304 calculates the actual maintenance cycle for each sensor based on the actual deterioration data calculated by the actual deterioration data calculation unit 303. This allows the maintenance cycle to be changed to suit the deterioration state of each sensor 21-23 after each maintenance. Furthermore, by comparing the actual maintenance cycles of the first sensor 21 and the second sensor 22 with the reference maintenance cycle, if the deterioration trends of the sensors are similar, the deterioration trends of the other sensors can be estimated. This allows the deterioration trends of sensors that cannot be stopped, such as sensors installed in the final discharge tank, to be estimated from the deterioration trends of other sensors in the same measurement system. As a result, the frequency and cost of maintenance for sensors that cannot be stopped can be reduced.

[0079] Other Embodiments The present invention is not limited to the above-described embodiments.

[0080] For example, the measurement system 100 may include a plurality of measuring devices 20. These measuring devices 20 may be arranged in different locations, such as different aeration tanks of a water treatment facility or different locations of a sample pipe. The information processing device 30 then performs an abnormality determination for each measuring device 20 in the same manner as in the above embodiment.

[0081] The measuring device according to the present invention is not limited to measuring liquid samples such as those used in water quality analysis as described above, but can also measure NOx, SO ２ , CO, CO ２ , O ２ The present invention can also be applied to measuring gas samples such as silica.

[0082] Some of the functions of the information processing device 30 described in the above embodiment may be provided in a computer or the like other than the information processing device 30. As an example, a function as a data storage unit that stores data such as output values ​​output from the sensors 21 to 23 may be provided in, for example, a cloud server or the like other than the information processing device 30.

[0083] In addition, the information processing device 30 may not have a machine learning unit, and a machine learning model that has been machine-learned on a computer or the like other than the information processing device 30 may be transmitted to the information processing device 30 via a communication line and implemented therein.

[0084] Furthermore, the present invention is not limited to the above-described embodiment, and it goes without saying that various modifications are possible without departing from the spirit of the present invention.

[0085] According to the present invention, in a measurement system that utilizes machine learning, it is possible to determine the need for maintenance of a measurement device or the need for re-learning of a machine learning model.

[0086] DESCRIPTION OF SYMBOLS 100: Measurement system 20: Measuring device 21: First sensor 22: Second sensor 30: Information processing device 31: Calculation unit 32: Display control unit 33: Memory unit 34: Machine learning unit 35: Prediction value estimation unit 36: Abnormality determination unit 301: Actual state measurement value calculation unit 302: Deterioration determination unit 303: Actual measurement deterioration data calculation unit 304: Maintenance cycle calculation unit

Claims

1. a measurement device having at least a first sensor and a second sensor; a predicted value estimation unit that estimates a predicted value of the second sensor from the actual measurement value of the first sensor using a machine learning model that has learned the relationship between the actual measurement value of the first sensor and the actual measurement value of the second sensor; and an abnormality determination unit that compares the predicted value of the second sensor with the actual measured value of the second sensor to determine whether an abnormality exists in the machine learning model or the measuring device.

2. 2. The measurement system according to claim 1, wherein the abnormality determination unit determines whether the difference or ratio between the predicted value of the second sensor and the actual measured value of the second sensor is within a predetermined tolerance range, and if the difference or ratio is outside the tolerance range, determines whether the abnormality exists in the machine learning model or the measuring device.

3. The measurement system according to claim 1 , wherein the abnormality determination unit determines whether an abnormality exists in the machine learning model or the measurement device using a non-parametric density estimation method.

4. 3. The measurement system according to claim 1, wherein the abnormality determination unit determines that the machine learning model is abnormal when the data density of the learning dataset from which the machine learning model was created is abnormal in a region where the data density is equal to or greater than a predetermined threshold, and determines that the machine learning model or the measuring device is abnormal when the data density is abnormal in a region where the data density is less than the threshold.

5. The measurement system according to claim 1 or 2, further comprising a machine learning unit that creates the machine learning model using a learning dataset consisting of the actual measurement values ​​of the first sensor and the actual measurement values ​​of the second sensor.

6. The measurement system according to claim 5 , wherein the machine learning unit re-learns the machine learning model when the abnormality determination unit determines that the machine learning model is abnormal.

7. The measurement system according to claim 1 or 2, further comprising: a maintenance signal that prompts a user to perform maintenance on the measurement device when the abnormality determination unit determines that the measurement device is abnormal.

8. 3. The measurement system according to claim 1, wherein each of the first sensor and the second sensor measures one of the conductivity, resistivity, oxidation-reduction potential (ORP), chemical oxygen demand (COD), turbidity, color, temperature, pressure, oil film, or concentration of a predetermined component contained in the sample.

9. a deterioration determination unit that determines whether a difference between the actual state measurement value and the predicted state value of the first sensor is within an allowable range, based on reference deterioration data that indicates a relationship between a predicted state value of the first sensor and time and actual deterioration data that indicates a relationship between an actual state measurement value of the first sensor and time, The measurement system according to claim 1 or 2, wherein when the deterioration determination unit determines that the state of the first sensor is outside the allowable range, an out-of-allowable range deterioration signal is output, indicating that the state of the first sensor is outside the allowable range of deterioration.

10. a measured deterioration data calculation unit that calculates first measured deterioration data indicating a relationship between the measured state value of the first sensor and time and second measured deterioration data indicating a relationship between the measured state value of the second sensor and time; The measurement system according to claim 9 , further comprising a maintenance cycle calculation unit that calculates a maintenance cycle for the first sensor and a maintenance cycle for the second sensor based on the first measured deterioration data and the second measured deterioration data.

11. 3. The measurement system according to claim 1, which is used for water quality analysis.

12. A method for determining an abnormality in a measurement system that measures a sample using at least a first sensor and a second sensor, comprising: estimating a predicted value of the second sensor from the actual measurement value of the first sensor using a machine learning model that has learned the relationship between the actual measurement value of the first sensor and the actual measurement value of the second sensor; A method for determining an abnormality in a measurement system, which compares the predicted value of the second sensor with the actual measured value of the second sensor to determine whether an abnormality exists in the machine learning model or the measurement system.

13. An abnormality determination program for a measurement system having at least a first sensor and a second sensor, a predicted value estimation unit that estimates a predicted value of the second sensor from the actual measurement value of the first sensor using a machine learning model that has learned the relationship between the actual measurement value of the first sensor and the actual measurement value of the second sensor; An abnormality determination program for a measurement system that causes a computer to function as an abnormality determination unit that compares the predicted value of the second sensor with the actual measured value of the second sensor and determines whether an abnormality exists in the machine learning model or the measurement system.