JP7119930B2 - Environment sensor detection data processing device, processing method, computer-readable storage medium, and environment sensor system - Google Patents

Description

The present invention relates to an environment sensor system, an environment sensor detection data processing apparatus, a processing method, and a computer-readable storage medium storing a computer program for causing a computer of the environment sensor to execute the method. , a plurality of environmental sensor devices (e.g., a PM sensor device, a sulfur oxide content sensor device, and a nitrogen oxide content sensor device) are arranged, and the environmental sensor devices are used to determine the content of various constituents or substances in the environment. Detecting quantities and providing desired detection accuracy, each environmental sensor device has an environmental sensor element or module and is connected to a network system.

Typically, typical environmental sensor systems for monitoring environmental conditions include atmospheric analyzers installed in locations such as cities and roads for government agencies to obtain environmental conditions in cities. The environmental sensor system measures environmental parameters such as environmentally sensitive gases contained in the atmosphere and publishes the results via the Internet. This is used to confirm the safety of the living environment, to formulate city and road plans by government agencies, and to subsidize environmental policies. At present, generally installed atmospheric analyzers are selected according to the environmental pollutants to be detected, and in some cases, a plurality of analyzers are combined into one analysis system. In recent years, for example, in developing countries, air pollution composed of particulate matter (hereafter referred to as PM) and harmful gases such as nitrogen oxides and sulfur oxides caused by the disposal of automobiles and factories are increasing. Emissions are a problem. Therefore, in order to monitor the state of air pollution, it is necessary to accurately detect the amount of PM contained in the atmosphere and the emission of other harmful gases.

Agencies use high-precision direct measurement equipment as baseline equipment, direct equipment as standards, and indirect measurement equipment to evaluate and select. For example, in the PM amount measurement method, the β-ray absorption method is a direct measurement device, and the Microbalance method is an indirect measurement device, and the correlation between the device and the β-ray absorption method is 0.98. and meet national standards, these two types of devices are widely used. For example, in the β-ray absorption method, atmospheric gas containing particulate matter is blown onto the filter material for one hour, the PM remaining in the filter material is measured, and the PM weight per unit volume of the air that has passed through is converted. Determine the concentration to obtain the abundance of PM in the atmosphere. Based on the measurement principle of the β-ray absorption method, it is necessary to accumulate a certain amount of PM in the filter material, so real-time measurement is difficult. For example, the Microbalance method indirectly measures the change in mass by measuring the vibrational frequency of a particular filter material, solving the problem of obtaining PM numbers in real time. The method is sensitive to environmental influences, requires strict exclusion of external influences and frequent maintenance of the equipment to ensure the accuracy of the measurements.

For example, the light scattering method is an indirect measurement method similar to the microbalance method, which solves the problem of the direct measurement method that PM values cannot be obtained in real time, and greatly reduces the influence of environmental factors on the measurement. However, it is not used by government agencies due to accuracy and cost issues. The light scattering method is one of the methods that can measure the amount of PM in the atmosphere in real time, and is the most common method for measuring the PM content in the atmosphere in the private sector. The principle is to irradiate light into the atmosphere of a constant blast, and measure the number of PM per unit volume of air that has passed, based on the number of times the light is blocked by the PM. The advantage of the light scattering method is that even if the measurement is performed in a short period of time, good measurement accuracy can be achieved, and real-time measurement can be achieved. However, in the β-ray absorption method, the weight is directly measured to calculate the PM concentration, but in the light scattering method, the PM concentration is estimated based on the numerical concentration. need to do In addition, since the detection accuracy of PM, which is a detection target, is likely to be interfered with by other factors, the accuracy may be poor.

Considering the influence of such environmental factors on the measurement accuracy, the environment around the mounting point may be changed, and for example, changes in the temperature, humidity, etc. around the mounting point may be maintained so that the surrounding environment does not fluctuate greatly. , for example an expensive temperature regulating Peltier module or the like. As a result, the temperature of the environment sensor element can be kept constant, but the cost of the device itself for maintaining the surrounding environment is very high, so the cost of the environment sensor device, which is a subunit, increases, and the cost of the entire device increases. rose significantly.

Therefore, in order to improve the measurement accuracy of the light scattering method, the following method of correcting the data measured by the light scattering method, correcting the quantity concentration data, and converting the quantity concentration to weight concentration with high accuracy is proposed. Proposed. For example, Patent Literature 1 discloses a method of correcting the quantity concentration and converting the quantity concentration into a weight concentration without changing the environmental factors. Not only PM analyzers but also other atmospheric analyzers have the same problem. Electronic elements themselves are susceptible to interference from environmental factors such as temperature and humidity, and electronic sensors for gases and substances other than those to be measured cannot be used. There are also cases of false reports and false positives. For example, when the object to be detected is CO, other hydrocarbons contained in the air, such as C in substances that are effective in gasoline, may interfere with the electronic sensor and affect detection accuracy. Therefore, it is necessary to reduce the interference of these interfering elements with the detection target.

In addition, the structure of the electronic environment sensor element is simple and the cost is low. Since the density can be greatly improved and the measurement of a fine area can be realized, the improvement of the accuracy of the sensor element instead of the expensive sensor device has become a focus of research in recent years. However, although the method and apparatus described in, for example, Patent Document 1 can improve the measurement accuracy, the cost of the apparatus also increases at the same time, and the cost advantage is not sufficient compared to the microbalance method. It is difficult to come up with a solution, and it is difficult to realize high-density installation.

Electronics can experience drift and measurement deviations after long periods of operation. Therefore, it is necessary to calibrate and compensate the environmental sensor after long periods of use. Similarly, the measurement accuracy after long-term use of other types of air analysis devices also has the same problem as the problem in Cited Document 1. That is, both electronic environmental sensors and large atmospheric analyzers have the following problems. When using these atmospheric analyzers, it is necessary to perform sufficient accuracy calibration in advance on the atmospheric analyzers, set each environmental sensor, and perform measurements using these environmental sensors. However, when using an air analysis device, the longer the measurement time is, the more susceptible it is to contamination, etc., and the analysis device may change over time, making it impossible to perform highly accurate measurements. When the preset maintenance period approaches, it is necessary to perform maintenance work such as cleaning, replacement, and recalibration of the environmental sensor element. Especially if there is an unspecified contamination situation in the detection environment, for example heavy contamination of several days or weeks, the maintenance cycle of the sensor will be significantly shortened, so that the data obtained will not be inaccurate. must be carried forward.

Moreover, the following method may be used as one of the methods for calibrating each environment sensor module. Prepare a plurality of known standard gas samples whose concentration is adjusted in advance for the substance to be detected, measure the concentration of these standard gases using the environmental sensor element to be calibrated, change the concentration of the standard gas sample, Obtaining a calibration curve showing changes in the value-measurement data, and calibrating the measurement module based on the curve to improve the accuracy of the measurements. Some environmental sensor modules can be calibrated in one setting, but due to variations in environmental sensor modules due to various external and internal factors, calibrate all environmental sensor modules individually when used for high-precision measurements. There is a need. Also, if multiple environmental sensor modules are installed in one environmental sensor device, each environmental sensor module must be calibrated one by one. As the number increases, the number of environmental sensor modules that need to be calibrated increases several times, greatly increasing the cost for calibrating the environmental sensors. In addition, after operating for a certain period of time after installation, it is necessary to periodically retrieve the environmental sensor device and check its measurement accuracy. If highly accurate data cannot be obtained, it is necessary to recalibrate and repeat the calibration operation. There is Even if the device to be calibrated is replaced with a calibrated one, it still needs to be adjusted in the field. In particular, most environmental sensors for measuring environmental information are installed in suburban areas where it is difficult to transport, and require high maintenance costs for recovery and resetting. Therefore, frequent calibration of environmental sensors inevitably leads to maintenance costs.

As in Patent Document 1 above, for a single sensor, data analysis can be performed on the temperature and humidity that cause interference to reduce the interference, but gases or substances other than the gas or substance to be measured can also be detected by the sensor. In the case of interference, the sensor cannot improve the measurement accuracy of the corresponding part. In particular, when objects that interfere with each other are collected at the same time, for example, when the contents of CO and hydrocarbons, which are C-containing compounds, are detected at the same time, only the temperature and humidity interference described in Patent Document 1 is corrected. methods to do so become inadequate.

Chinese Patent Application Publication No. 105300861

INDUSTRIAL APPLICABILITY The present invention provides an environment sensor detection data processing apparatus, processing method, and computer reading that can improve the accuracy of the data of the environment sensor device at low cost and reduce the installation and maintenance costs of the environment sensor system. It is an object of the present invention to provide a possible storage medium and environmental sensor system. Another object of the present invention is to provide a sensor device capable of improving the resistance to interference. It is another object of the present invention to provide an environment sensor device capable of extending the maintenance period.

In one aspect of the present invention, an environment sensor device is a detection data processing device for processing measurement data obtained by measuring the concentration of a detection target in an environment, wherein in a learning mode, the environment sensor device a measurement data input unit that acquires a plurality of learning measurement data and acquires the actual measurement data of the environment sensor device in an actual measurement mode; a true value acquisition unit that acquires true values corresponding to the learning measurement data; A plurality of association data are obtained by associating a plurality of learning measurement data with the corresponding true values, and some association data are randomly selected from the plurality of association data to create an association data set for learning. and repeatedly executing the process of processing the association data set using the neuron network model to obtain the correction characteristics a plurality of times, obtaining the plurality of correction characteristics, and executing the processing of obtaining the correction characteristics for each of the correction characteristics. a correction characteristic generator that verifies the correction characteristics based on the true values in the association data that are not used when performing the correction, and selects the optimum correction characteristics based on the verification results; and the actual measurement based on the optimum correction characteristics and a measurement data corrector for correcting the data.

According to this aspect, the device for processing data detected by the environment sensor acquires, from the environment sensor, the plurality of learning measurement data of the environment sensor device in the learning mode and the actual measurement data of the environment sensor device in the actual measurement mode by the measurement data input unit. Then, the true value acquisition unit acquires the true values corresponding to the learning measurement data from the environment sensor device, and the correction characteristic generation unit associates the plurality of learning measurement data with the corresponding true values to produce a plurality of Acquiring association data, randomly selecting some association data from a plurality of association data to create an association data set for learning, and processing the association data set by a neuron network model to obtain correction characteristics The process is repeatedly executed a plurality of times to obtain a plurality of correction features, and for each correction characteristic, the correction characteristic is obtained based on the true value in the association data that is not used when executing the process for obtaining the correction characteristic. After verification, the optimum correction characteristic is selected based on the verification result, and the measured data is corrected based on the optimum correction characteristic by the measurement data correction unit. According to the processing device, regardless of the accuracy of the environmental sensor, that is, whether the environmental sensor has a relatively high accuracy or the environmental sensor has a relatively low accuracy, the true value obtained from the environmental sensor is obtained by the neuron network. Based on the training data, a model of the neuron network can be learned, and based on the obtained model of the neuron network, the measured data obtained from the environmental sensors can be corrected. In addition, a process of randomly selecting a part of association data from a plurality of association data to create an association data set for learning, and processing the association data set by a neuron network model to obtain a correction characteristic is repeated multiple times. While executing repeatedly, each correction characteristic is verified based on the true value in the association data that is not used when executing the process of acquiring the correction characteristic, and the optimum correction characteristic is obtained based on the verification result. By selecting , the existing data can be used repeatedly with a small amount of data, and the calibration can be repeatedly performed using the true value. Waiting time can be reduced. In addition, since the accuracy required for the sensor itself is also reduced, a relatively inexpensive sensor element can be used to achieve a data processing effect equivalent to that of an expensive high-precision sensor. Therefore, the objective of cost reduction can be achieved. Moreover, since the learning function of the neuron network is used, it is possible to achieve higher data conversion accuracy using less original data. In addition, the learning capabilities of neuron networks can reduce the cost of manual data statistics and analysis and manual model building by humans, and high data accuracy can be achieved by matrix algorithms. Even if the data detected by the environmental sensor device deviates from the true data for some reason, it is possible to create corrected data using processing based on the neuron network model and obtain data that is close to the true data. .

This aspect further includes an operation state information input unit for inputting operation state information of the environment sensor device, wherein the operation state information includes internal and/or external temperature, humidity, At least one of atmospheric pressure, degree of illumination, ID information of the environment sensor device, and geographic location information of the environment sensor device is included, and when the association data is acquired, the operation state information is further added for association. .

By using the above operating state parameters, data can be detected based on these parameters and true values, and a model of the neuron network can be quickly generated, further shortening the time required to put the environmental sensor device into practical use. . In addition, since the accuracy of these acquired information was ensured at the time the environmental sensor device was installed, the usage time of the device can be predicted by artificial intelligence analysis of the time parameter, and the length of the operating state information of the environmental sensor device can be estimated. We can ensure the provision of the period. Therefore, according to this aspect, accurate detection data can be provided for specific gases.

In this aspect, the detection target is a PM value and/or a gas component, and the gas component contains at least one of nitrogen dioxide, sulfur dioxide, carbon monoxide, ozone, carbon dioxide, hydrogen sulfide, and VOC. include.

As a result, the gas component can be detected, and the concentration of the environmental pollutant gas can be detected.

In this aspect, the processing device uses a neuron network model based on measurement data obtained by a plurality of types of sensor devices for different detection targets, and adds mutual interference characteristics between different detection targets to the correction features. .

As a result, the concentrations of multiple types of detection target components that interfere with each other can be calculated, and the ability and level of mutual interference between multiple types of detection targets having mutual interference characteristics are used to create correction features having mutual interference characteristics. This can be used to correct the actually measured data of the object to be detected, and more accurate data can be obtained.

Alternatively, the plurality of learning measurement data and actual measurement data acquired from the environment sensor device by the measurement data input unit are learning measurement data and actual measurement data corrected in advance based on the operating state information by the sensor device. Alternatively, after the measurement data input unit acquires a plurality of pieces of learning measurement data and actual measurement data from the environment sensor device, the data pre-correction unit corrects the learning measurement data and the actual measurement data based on the operating state information. It is corrected in advance and supplied to the correction characteristic generator. The pre-correction may be performed using a neuron network model, similar to the generation of correction characteristics by the correction characteristics generation unit.

By this, before formally generating the correction data using the correction characteristic generation unit, the learning measurement data and the actual measurement data are pre-corrected once based on the operating state information. By performing the correction twice, it is possible to eliminate a large deviation between the data before pre-processing and the correct data, thereby improving the accuracy of the correction by the correction characteristic generator. In addition, when performing the second correction using the correction characteristic generation unit, the mutual interference characteristic is also taken into consideration when generating the correction characteristic, thereby reducing interference between multiple types of detection targets that have mutual interference. can be made

In this aspect, the true value is at least one of data acquired by a verified environment sensor device, true data stored on the Internet, and data detected by a high-precision analysis device.

This allows the true value to be obtained using a variety of methods, reduces the demands on the data source of the processor, and reduces the cost.

In this aspect, when the measurement data input unit and the true value acquisition unit acquire the corresponding measurement data, the operation state information input unit further acquires time point data of the measurement data, and acquires the association data. At this time, the time point data is further added to the operation state information and associated.

As a result, since the operating state information includes the time parameter, the detection data of the sensor can be corrected based on the time parameter, so that the maintenance period of the environment sensor system can be extended. That is, according to this aspect, the learning function of the neuron network analyzes the time in the operating state information of the environment sensor, estimates the attenuation of the device and the state of change over time, and thereby improves the measurement accuracy of the device within a longer period of time. can be ensured, the frequency of calibration of the device can be reduced, and the maintenance cost can be reduced. Other neuron networks can analyze big data and realize the origin analysis and prediction of environmental data. In this way, the period of use before calibration and maintenance of the environment sensor device and the environment sensor system can be extended.

For example, as a correction processing model, a correction method using a neuron network (NN) with a two-layer tree structure is adopted.

The first-layer neuron network performs one-time correction for each environmental sensor element or module obtained. The one-time correction may use a standard neuron network, a standard neuron network modified according to a conventional curve formula, or directly using a conventional curve formula without using a neuron network. good too. The input data are output data of the environmental sensor element or module and operating state information of the environmental sensor.

By reducing or eliminating hidden layers, the neuron network of the first layer preferentially eliminates the influence of the operating state of the environmental sensors on the output data, and avoids the local optimum of vanishing gradients. In this embodiment, we use one hidden layer, the standard function of X*W1+b1. Independent correction provided more flexible combination options with environmental sensor elements or modules that eliminated the effects of existing correction formulas or operating conditions.

The input of the neuron network of the second layer is the detection data and operating state information of all environmental sensor elements or modules, and the output is the corrected data of all environmental sensor elements or modules. In this embodiment, we use one hidden layer, the standard function of X*W1+b1. More hidden layers may be used depending on the computing power of the device, and the function or neuron network itself may be modified as needed.

As a specific example, in the data processing using the neuron network model, the data detected by the environmental sensor element or module and the operating state information of the environmental sensor device are constructed into one data pair, and the data pair and the Aggregate the true values, introduce the aggregated data into a neuron network artificial intelligence training system, train feature samples of the data, and obtain a model. Here, the environmental sensor element or module has the function of measuring particulate matter present in the atmosphere and has the function of detecting the concentration of gases that affect the environment. Alternatively, the environmental sensor element or module has a sub-element or sub-module for measuring particulate matter present in the atmosphere and has a sub-element or sub-module for detecting the concentration of gases affecting the environment. In this way, a function-integrated environmental sensor element or module may be used, or sub-elements or sub-modules each implementing multiple functions may be used.

Accordingly, the above system performs model-based correction processing based on these data, and corrects the measured data to the true value or data close to the true value to the server device. Further, after the data is stored in the server, the server may provide a data search service, a data graphic service, and a data analysis service. Data required by the operator may be displayed on the display based on the operator's request or command.

Data processing is not limited to correction processing, and may further include conversion processing between different types of data, for example, processing for converting particle count data into weight data, or conversion processing between data with different specifications. Further, the correction is not limited to correcting data with deviation to a true value or data close to the true value, and further includes data format, specification correction, and the like.

Also, when constructing a model of a neuron network, as an example, the total number of pieces of association data between important data and true values used is about 1000. By using about 1000 associated data, deviation data can be corrected to accurate data using a neuron network model based on a large amount of data. In addition, in this aspect, by using a two-layer neuron network, it is only necessary to use about 1000 association data, the requirement for the amount of association data is low, the amount of data processing is reduced, and the construction of a neuron network model Time can be shortened, and the time required from the installation of the environment sensor system of the present invention to its own correction can be shortened.

The larger the number of data for learning a neuron network model, the higher the accuracy of correction. For example, in one aspect of the present invention, in order to obtain sufficient correction results and reduce the time required for data collection, training data and measurement data are randomly grouped multiple times, and the data is Use multiple times. Specifically, in the case of 1000 data sets, 900 data are randomly extracted as learning data, 100 data are used as measurement data, and 100 models (actual processing ), the average value of the obtained models is obtained, and the result of about 1000×100 data sets can be obtained. With such a processing method, the characteristics of the neuron network analysis method can be used well, and the original data accumulation time and initial setting maintenance time can be greatly reduced. In addition, compared to conventional data analysis of non-neuronal networks, the data utilization rate increases geometrically, achieving higher data conversion accuracy using less data, and reducing the dependence on the original data. can be done.

If the true value of the data can be obtained by a network or other calibrated environmental sensor device near the installation point, there is no need to perform calibration before installation, and the cyclic neuron network learns the true value of the vicinity, Outputting predicted operating state information of the uncorrected environmental sensor device, predicting the true value using the neuron network based on the data output by the calibrated environmental sensor device, and not calibrating before installation Correct the inaccurate data acquired by the environment sensor device to the predicted true value. As the accumulated measurement data increases, the increase in the learning data is also reflected in the learning of the model, so the deviation between the true value and the measurement data is gradually corrected by the model.

The operating state information of the environment sensor may further include a time parameter, and time is also one of the important conditions affecting the true value of data. For example, by learning data collected over a month, a neuron network can detect different degrees of data drift and decay for each environmental sensor over a month and obtain a model of time and decay. As another example, regular maintenance of the instrument allows data to be obtained over a longer period of time, allowing the drift and decay models to be more accurately predicted. The model is the model of the device with the same model number, and the accuracy of the device with the same model number can be improved. For this reason, even if there is a deviation between the measured data and the true value due to changes in the environmental sensor elements in each environmental sensor device over time during long-term measurement after installation, the model is trained using the measured data accumulated in advance. , changes over time can be learned, the maintenance period can be extended, and the frequency of recalibration can be reduced.

Furthermore, the sensor device is a sensor device in which some or all of the sensor elements are not calibrated. In addition, in the present invention, even if the sensor device used is a sensor with normal accuracy or relatively low accuracy, the model of the neuron network using the true value can be used to obtain high-standard, high-precision data obtained by the sensor device. Detected data that does not meet the requirements of the standard can be converted into detected data that meets the requirements of high standard and high accuracy. Therefore, the cost of the entire sensor system can be reduced.

In another aspect of the present invention, a plurality of sensor devices connected to a network and the processing device described above connected to the network are included, and the measurement data input unit of the processing device inputs via the network and acquiring actual measurement data of the plurality of sensors, and correcting the actual measurement data.

This makes it possible to provide a sensor system with all the advantages and advantages of each of the processors described above.

In this aspect, the sensor device is a sensor device in which some or all of the sensor elements are not corrected, and/or the sensor device includes a sensor that measures concentrations of a plurality of types of detection targets, and an environment sensor. and an operating state information acquiring unit group that acquires operating state information of the device with high precision.

According to another aspect of the present invention, there is provided a method for processing detection data of an environment sensor for processing measurement data obtained by an environment sensor device measuring the concentration of a detection target in an environment, the method comprising: A measurement data input step of acquiring a plurality of data for learning measurement of a sensor device, acquiring actual measurement data of the environment sensor device in an actual measurement mode, and a true value acquisition step of acquiring true values respectively corresponding to the data for learning measurement. and obtaining a plurality of association data by associating the plurality of learning measurement data with the corresponding true values, and randomly selecting a portion of the association data from the plurality of association data to obtain an association data set for learning and obtain correction characteristics by processing the association data set using a neuron network model to obtain a correction characteristic a plurality of times, obtaining a plurality of correction characteristics, and obtaining a correction characteristic for each of the correction characteristics. a correction characteristic generation step of verifying each correction characteristic based on a true value in association data that is not used when executing the process and selecting an optimum correction characteristic based on the verification result; and a measurement data correcting step of correcting the actual measurement data by using an environment sensor.

In another aspect of the present invention, there is provided a computer-readable storage medium storing a computer program, wherein when a processor included in the environment sensor system executes the computer program, the environment sensor system a measurement data input step of acquiring a plurality of learning measurement data of the environment sensor device in a learning mode and acquiring actual measurement data of the environment sensor device in an actual measurement mode; acquiring a plurality of association data by associating a true value acquisition step to be acquired with the plurality of learning measurement data and the corresponding true values, and randomly selecting a portion of the association data from the plurality of association data; A process of creating an association data set for learning and processing the association data set by a neuron network model to obtain correction characteristics is repeatedly executed a plurality of times to obtain a plurality of correction characteristics, and for each of the correction characteristics a correction characteristic generation step of verifying each correction characteristic based on a true value in association data that is not used when executing the process of acquiring the correction characteristic, and selecting the optimum correction characteristic based on the verification result; and a measurement data correction step of correcting the measured data based on the optimum correction characteristics.

The present invention can reduce the cost of manual model building, reduce the cost of pre-correction of the environmental sensor device before installation, and reduce the deviation of the measurement data due to the aging of the environmental sensor device. Provided is an environment sensor detection data processing device, a processing method, a computer-readable storage medium, and an environment sensor system that can reduce the frequency of calibration, lengthen the maintenance cycle, and reduce the maintenance cost of the environment sensor device. can.

1 is a diagram showing an environment sensor system according to a first embodiment; FIG. FIG. 4 is a diagram showing the configuration of a neuron network implementation model used in the processing of the server device of the environment sensor system; It is a detailed diagram of the construction and implementation of the first neuron network (NN) in the diagram of the tree structure of the implementation model. FIG. 2 is a detailed diagram of the construction and implementation of a neuron network (NN) for the second time in the diagram of the tree structure of the implementation model; 4 is a table showing the state of interference between gases; FIG. 4 is a diagram showing a data processing method of the present invention; It is a figure for demonstrating the effect obtained by the environment sensor system of this invention. It is a figure which shows the environment sensor system which concerns on 2nd Embodiment. It is a figure which shows the environment sensor system which concerns on 3rd Embodiment.

The following describes preferred embodiments of the invention.

[First embodiment]
FIG. 1 is a diagram showing an environment sensor system according to the first embodiment. A plurality of atmospheric analysis environment sensor devices 12 are installed. Inside the atmosphere analysis environment sensor device 12, a light scattering type PM environment sensor element, nitrogen oxides, sulfur oxides, carbon monoxide, ozone, carbon dioxide, and sulfuric acid are contained in one housing (not shown). Environmental sensor elements capable of detecting the concentration of gases that affect the environment, such as hydrogen and VOCs, are mounted, and constitute a group of environmental sensor elements for target substances to be measured. Here, an example of detecting the concentration of a plurality of types of gases to be detected with one environment sensor element is shown, but as another aspect, one environment sensor element may be provided for each gas to be detected. Alternatively, a plurality of environment sensor elements may be used in combination for a plurality of types of gases to be detected. Alternatively, one environment sensor element may be provided for some detection target gases, and one or a plurality of environment sensor elements may be provided individually or collectively for other detection target gases.

Overall, the analytical environmental sensor device 12 may also measure the concentration of particulate matter present in the atmosphere, i.e. PM concentration, and the concentration of gases affecting the environment, i.e. nitrogen oxides, Concentrations of sulfur oxides, carbon monoxide, ozone, carbon dioxide, hydrogen sulfide, and VOCs may be measured. As an example, the analysis environment sensor device 12 may transmit to the wireless base station 13 the PM concentration data 12a measured for a predetermined period, for example, every minute. In addition, the atmospheric analysis environment sensor device 12 stores ID information indicating from which environment sensor device the measurement information is acquired in the same housing (not shown) as the environment sensor for the target substance to be measured. A device to acquire, a device to acquire positional information (for example, GPS) of the installation position, a group of devices such as a thermometer, a hygrometer, a barometer, and an illumination intensity meter are further mounted, and acquire the data 12b of the operating state information of the sensor. device. Acquire operating status information of the environment sensor device before installing it in the environment sensor device. These devices ensure general accuracy and can collect information when maintenance is not required for a long period of time. is. As another example of the present invention, these operating state information may be corrected as information to be corrected. For the correction method in this case, refer to the method of correcting detection data of the detection target gas of the present invention. good too.

In the same housing, a battery (power supply) for driving these environment sensor elements, a memory card for data storage, a communication device for transmitting measurement data to the server device 15 via the Internet 14, and an environmental sensor element are installed. Ancillary equipment (not shown), such as equipment that maintains the operating environment, are also included. These auxiliary devices are not essential devices. For example, in the present invention, a battery may be used as a power supply device. Also, the storage device may be a non-volatile memory card or a volatile storage device such as a flash memory. Also, the transmission device is not particularly limited, and any device capable of realizing data transmission and transmission may be used.

The communication system may use a wired system and/or a wireless system, but the wireless system is preferable because it can improve the degree of freedom in the installation position of the environment sensor device. Any type of wireless system may be used as long as data can be transmitted to a server device on the Internet. In this embodiment, a mobile communication system such as a 4G or next-generation 5G communication system may be used, a wide area wireless Internet such as WIFI may be used, or a short-range wireless network such as Bluetooth (registered trademark) may be used. may be used.

In this embodiment, the measurement data 12a and the operating state information 12b of each environmental sensor element are transmitted to the server device 15 via the Internet 14 by a wireless communication system, and stored in a storage device (not shown) in the server device 15. , is accumulated.

In this embodiment, the server device 15 processes the accumulated data. Further, in this embodiment, the data 16 published on the network is used as the necessary true value data in the learning data. Measurement data 16a of PM, nitrogen oxides, sulfur oxides, carbon monoxide, ozone, carbon dioxide, hydrogen sulfide, VOC, etc. measured by a highly accurate atmospheric analyzer, and the true value data Data 16b indicating operational state information (eg position and time) is used. The time parameters in the measurement, that is, the measurement data 12a transmitted from the small air analysis environment sensor at the same time, the operating state information 12b when detecting the above data acquired by the small air analysis environment sensor, published on the network High-precision density data 16a, which is the true value obtained by the process, and operation state information 16b at the time of acquiring the true value data are collected and stored in a data set 17. FIG. As shown in FIG. 1, in the specific processing method used in the present invention, measurement data 12a and operating state information 12b are associated with data (in this embodiment, data pairs are used; The concentration data 16a and the operating state information 16b are created as association data (similar to the above), and the created association data is sent to the server device 15. Store and accumulate in memory.

In the server, these accumulated data sets 17 are used as learning data, and the model 18 is made to learn. In this embodiment, a neuron network model indicated by reference numeral 12 in FIG. 2 is used as the model. The more data you learn, the better your model learns.

One neuron network model may be prepared for each environmental sensor. For example, for an environment sensor device equipped with 10 kinds of environment sensor elements, 10 neuron network models corresponding to each environment sensor are prepared. The number of layers of the neuron network and the correlation coefficient may differ according to each environmental sensor.

Depending on the type of environment sensor, if there are about 1000 or more pieces of learning data, model learning can achieve sufficient accuracy. After model learning is completed, each measurement data 12a transmitted from each uncalibrated environment sensor device is corrected using the model, and highly accurate corrected data 18a is obtained. Further, the data after correction may be displayed on the display device according to user's demand or request.

At present, sensor manufacturers perform different degrees of data correction on their sensor elements and use temperature and humidity compensation formulas or neuron network technology, but usually high-precision environmental sensor elements such as temperature and humidity The price of the operating state sensor is higher than the environmental sensor element itself, and the low accuracy seriously affects the measurement accuracy, and the high accuracy requirement cannot be fully achieved. Therefore, it is practically impossible to use the method of controlling the operating environment of the sensor with an expensive device. It obviously increases the cost of the equipment and makes the product lose market competitiveness. In the present invention, on the other hand, multiple environmental sensor elements share a single set of highly accurate operating state devices, thereby significantly reducing the cost of the device.

Another point of the present invention is to solve the problem of mutual interference of measurement data between a plurality of sensor elements. At present, gas sensors are usually sensitive to a plurality of gases to be measured, especially combustible or oxidizing gases. For example, as shown in FIG. 5A, when measuring a standard gas in a laboratory, input a standard gas containing only 100 ppm nitrogen dioxide, which was erroneously measured by other non-nitrogen dioxide sensors. value appears. When a standard gas of 200 ppm nitrogen dioxide is input, the non-nitrogen dioxide sensor also fluctuates, as shown in FIG. happens. A standard gas may be repeatedly measured and the variation curve of the false measurement sensitivity of each sensor may be calculated. increase. Furthermore, in the actual state where multiple types of gases coexist, it is not possible to determine the sensitivity of erroneous measurement in the output data and the ratio of actual measurement data in the measurement results shown in FIG. It cannot be guaranteed to estimate the variation rule of sensitivity to obtain an accurate figure. On the other hand, the method of the present invention can avoid the inaccuracy of the setting formula after processing with the model of the neuron network, and the processing with the alternative model of the neuron network improves the accuracy.

That is, the environmental sensor device is provided with a sensor element for each gas, but conventional gas sensors are sensitive to a plurality of types of gases. For example, the following shows the sensitivities of multiple types of environmental sensors to target gases. Specifically, when the target gas to be measured is measured, not only the target gas but also one or more kinds of interfering gases coexist in the measurement environment. It is the sum of the concentration measurement result and the interference value of the interfering gas concentration (also referred to herein as sensitivity). Depending on the target gas, different kinds of interference gases and different gas concentrations affect the interference value.

In contrast, the environmental sensor system of the present invention is used to perform corrective measurements, obtain multiple environmental sensor measurements simultaneously, and process the multiple measurements based on the corresponding true values by a neuron network model. Then, the interference value formed by the interfering gas is removed from the obtained measurement data to obtain the correction characteristic of the measurement data. That is, the difference between the obtained measurement data and the true value can be made smaller than the predetermined threshold by correcting with the correction characteristic. Then, when actually measuring, it is possible to correct the actual measurement data using the obtained correction characteristics and obtain accurate data.

For a specific data processing method, an example of the data processing method shown in FIG. 6A may be referred to. Here, a plurality of sets of operating state information of the sensor device at the same point in time, sensor detection data, and true values are used as one data set to obtain measurement data 51, and a neuron network having the configuration of FIG. 2 corresponding to each sensor is constructed. Input the same data to the model 18 and learn the data. The neuron network performs data processing comparison on the input operating states, sensor detection data and true values at the same time, extracts and stores characteristic data. Detected data 52 is input to a model trained on sufficient data, the true value of the data corresponding to each sensor is estimated from the data processing characteristics stored in each model, and processed data 53 is output.

The learning function of the neuron network may reacquire the rules between the data by increasing the hidden layer and changing the dimension of the data, and the realization of this part is described in FIG. The overall tree structure of the neuron network is described.

FIG. 2 is a diagram showing the configuration of an implementation model of a neuron network used in the processing of the server device of the environment sensor system. As an example, in the configuration diagram shown in FIG. 2, the input data is the output data 12a, 12b or 82a, 82b of the atmospheric analysis environment sensor device, the neuron network model for processing is 18, and 24an is the nth represents the model used by the first layer neuron network 22 for the first correction of the output data of the environmental sensor element or module, 24b represents the model used by the second layer neuron network 23, the output data is 18a, Here, 23an represents the corrected numerical value of the output data of the n-th environmental sensor element or module. The initial correction here corresponds to the pre-correction in the aspect of the present invention. Further, the pre-correction of the data using the neuron network model by the server device is merely an example. The correction may be performed using the operating state information based on. Alternatively, it may be pre-corrected data corrected by a method such as manual calculation by the sensor device itself.

This embodiment employs a correction method using a mixed neuron network (NN) model with a two-layer tree structure. The first-layer neuron network performs one-time correction for each environmental sensor element or module obtained. The one-time correction may use a standard neuron network, a standard neuron network modified according to a conventional curve formula, or directly using a conventional curve formula without using a neuron network. good too. That is, the input data to be input to the neuron network are the output data of the environment sensor element or module and the operating state information of the environment sensor. By reducing or eliminating hidden layers, the neuron network of the first layer preferentially eliminates the influence of the operating state of the environmental sensors on the output data, and avoids the local optimum of vanishing gradients. Independent correction provided more flexible combination options with environmental sensor elements or modules that eliminated the effects of existing correction formulas or operating conditions. The input of the neuron network of the second layer is the detection data and operating state information of all environmental sensor elements or modules, and the output is the corrected data of all environmental sensor elements or modules. In this aspect, the implemented neuron network uses all one hidden layer, the sigmoid activation function. Also, more hidden layers may be used depending on the computing power of the device, and the function or neuron network itself may be modified as needed.

FIG. 3 is a detailed diagram of the construction and implementation of the first neuron network (NN). The implementation of this part uses a standard neuron network, uses the python language tensorflow to build a network model, and the algorithm and principle may refer to "Machine Learning" Chapter 5 Neuron Network of Tsinghua University Press . In the implementation, one hidden layer is provided, also called a two-layer neuron network, the hidden layer has 10 nodes, represented by n nodes in the detailed diagram, using the sigmoid function as the activation function, the cost The calculation uses the square of the difference between the true value and the output value. As shown in FIG. 3, the neuron network is divided into three layers: input layer, hidden layer and output layer, in this implementation one hidden layer is used, also called two-layer neuron network, the reason for the choice is This is because it has been theoretically proven that a two-layer neuron network can infinitely approximate arbitrary continuous functions. Those within the circular nodes represent the data and the lines of arrows are referred to as weights or weighted values W.

The calculation formula is as follows.

h _t =σ(W _h [h _t−1 , x _t ]+b ₁ )
O _t =σ(W _o [h _t−1 , x _t ]+b ₂ )
For example, an example calculation of the hidden layer is as follows.

h1=sigmoid(12an*Wanh1+12b1h1*Wb1h1+...+1*W1h1)
Similarly, the hidden layers h1-hn can be calculated based on the known W1 matrix, 22an can be calculated based on the W2 matrix, and the corrected data O _t can be obtained, and the actual calculation is performed by the corresponding neuron network program. This embodiment uses the python language tensorflow. On the other hand, when the output data of the environment sensor device and the true value 16an are acquired, the learning program of the neuron network follows the penalty mechanism until the matrix closest to the output values 22an and 16an is acquired. Adjust the numerical value of the data in W based on the squared difference of . The more training data, the more accurate the obtained W matrix. Although this embodiment uses a two-layer neuron network, the number of hidden layers is not limited, and other neuron network models, other modifications, and other activation functions may be used.

FIG. 4 is a detailed diagram of the construction and implementation of the second neuron network (NN) in the tree diagram. The structure and algorithm of the neuron network used in this embodiment are the same as those in FIG. Output all corrected data. The input data of the neuron network in the second layer of the tree structure has general accuracy, can remove the effects of working conditions, and deeply analyze the effects between each environmental sensor element or module. The method is not limited to standard neuron networks, but in practical use, the number of hidden layers can be increased according to the server performance, and modified accordingly to eliminate gradient vanishing, so as to achieve higher data accuracy. Local optima may be avoided. Variants of cyclic neuron networks or related network simulations may be applied to broadly extend practical applications such as data-driven prediction.

FIG. 6B is a diagram for explaining the effects of the present invention. As shown in FIG. 6B, the correlation coefficient between the uncorrected measurement data and the true value is adopted as the evaluation standard. As shown in FIG. 6B, in the correction of the PM2.5 environment sensor element, the correlation coefficient between the uncorrected measurement data and the true value is 0.86, while the correlation coefficient obtained by the application model processing method according to the present invention is The correlation coefficient between the corrected measured data and the true value is 0.95. Therefore, correction can be performed with high accuracy.

As a result, in the environment sensor system of the present invention, by using a neuron network model, it is possible to reduce the cost of manually constructing a model and reduce the cost of pre-correction of the environment sensor device before installation. In addition, when deviation occurs due to interference in the data output by the environmental sensor, the data output by the environmental sensor element can be calibrated and corrected using the neuron network model by spontaneously using the accumulated data. Therefore, it is possible to reduce the deviation of the measurement data due to the aging of the environment sensor device. Therefore, even if a deviation occurs in the environment sensor device, by correcting the output data using the neuron network model, the frequency of recalibration can be reduced, the maintenance cycle can be lengthened, and the environment sensor device Maintenance costs can be reduced.

The data processing method described above may also be implemented by computer software, and the software is stored in a computer-readable medium, such as a read-only storage medium such as a CDROM, DVDROM, or the like. Alternatively, a rewritable fixed or insertable storage medium such as an HDD or SD memory may be used. Alternatively, a storage medium such as a network disk on the network may be used. A computer program is stored in the storage medium storing the software, and causes the environment sensor system to implement the following steps when causing a processor included in the environment sensor system to execute the computer program. In an environmental sensor system composed of one or more environmental sensor devices and a server device connected via the Internet, the environmental sensor device can measure the substance content detected by one or more built-in environmental sensor elements or modules. Detected data and operating state information of the environmental sensor devices at the same time are collected at the same time, and the server device of the environmental sensor system receives the detected data detected by the environmental sensor elements or modules of one or more environmental sensor devices via the Internet. and the operating state information of the environmental sensor device, accumulates the detection data and the operating state information, and the server device stores the data detected by the environmental sensor element or module of the environmental sensor device and the operating state information of the corresponding environmental sensor device. to generate association data, input the association data into a neuron network model, and perform data processing including at least correction on the data detected by each environmental sensor element or module by data processing of the model, Output the corrected result data. Here, the neuron network model is constructed by training the training data.

Further, in this embodiment, about 1000 pieces of association data are used when creating a model. The reason is as follows. Conventional sensor devices typically calibrate a single sensor device or sensor elements within a sensor device individually. Even if the detection of the device itself is accurate, the final deviations appear in the data output by the sensor device. On the other hand, in this embodiment, by using about 1000 pieces of association data, deviation data can be corrected to accurate data using a neuron network model based on a large amount of data. In addition, in this embodiment, by using a two-layer neuron network, it is sufficient to use about 1000 pieces of association data. The construction time can be shortened, and the time required from the installation of the environment sensor system of the present invention to its own correction can be shortened.

Furthermore, according to the present invention, even if the sensor device used is a normal-accuracy sensor or a low-accuracy sensor, the neuron network model uses the true value to obtain a high-standard, high-precision sensor device acquired by the sensor device. The sensing data that does not meet the requirements can be converted into the sensing data that meets the requirements of high standard and high accuracy. Therefore, the cost of the entire sensor system can be reduced.

Further, in the present embodiment, in the server device 15, the means (unit or module) for acquiring measurement data from the sensor device corresponds to the measurement data input unit of the present invention, and acquires the true value from the true value data published on the network. The acquisition means (unit or module) corresponds to the true value acquisition section of the present invention, and the means (unit or module) to process the association data by the neuron network model and acquire the correction feature is the correction characteristic generation section of the present invention. and the means (unit or module) for correcting the measured data based on the optimum correction characteristic selected by the correction characteristic generation section corresponds to the measurement data correction section of the present invention, and acquires each operating state information from the sensor device. The means (unit or module) for doing so corresponds to the operation state information input section of the present invention.

Also, the software itself may be created using a conventional programming language such as JAVA (registered trademark). Since the execution method of the program is the same as or basically the same as the operation method of the environmental sensor correction method, detailed description thereof will be omitted here.

[Second embodiment]
FIG. 7 is a diagram showing an environment sensor system according to the second embodiment. In the correction method of this embodiment, unlike the first embodiment in which the data of the analysis device published on the network is used as the true value data, the atmospheric analysis environment that can realize pre-calibrated high-precision measurement A sensor device is prepared, and measurement data obtained from the environmental sensor is taken as a true value. An environment sensor that may be calibrated in advance using an environment sensor device having the same structure as an uncalibrated environment sensor device, and that can detect the amount of the same substance as the environment sensor element built in the environment sensor device. If so, environmental sensors with other structures may be used. The same constituent elements in the drawings of the first embodiment are denoted by the same reference numerals, and descriptions thereof are omitted.

A plurality of uncalibrated small atmospheric analysis environment sensor devices 12 are installed. In the vicinity of one of the environmental sensor devices 12, a proximate atmospheric analysis environmental sensor device 72 is provided for obtaining calibrated data of the same type of environmental sensor device structure capable of achieving pre-calibrated high-precision measurement. there is The communication system in this embodiment employs a wired connection method (a wireless connection may be used as necessary). The measurement data 12a, 72a and the operating state information 12b acquired by each atmospheric analysis environment sensor device 12 are transmitted to the server device 15 via the Internet 14 and stored in a storage device (not shown) of the server device 15. there is

In this embodiment, as true value data necessary for data learning, measurement data 12a and 72a transmitted from a small atmospheric analysis environment sensor and operating state information are aggregated at the same point in time as measurement, and a data set 17 is obtained. Remember. In the server device 15, these accumulated data sets are used as learning data to make the neuron network model indicated by reference numeral 12 in FIG. 2 learn. After model learning is completed, each measurement data 12a and operating state information 12b transmitted from each uncalibrated environment sensor device are corrected using the model 18, and highly accurate corrected data 18a are obtained and corrected. The later data 18a is used as corrected data transmitted by each environment sensor device after correction. For example, the corrected data is provided to an information gathering device or the public, or displayed on a display device or the like.

The correction method may use the correction method described in the first embodiment, or may be realized by a storage medium storing the software described in the first embodiment.

[Third Embodiment]
FIG. 8 is a diagram showing an environment sensor system according to the third embodiment. In the present embodiment, unlike the first embodiment, in which analytical device data published on the Internet is used as the true value, there is an atmospheric analysis device whose structure should be corrected in the vicinity of the analytical device 89 that ensures high accuracy. A device 82 that is the same as the environment sensor device 12 is provided and simultaneously measures data 82c of the same atmosphere at the same time and information 82d of operating state information. Parts having the same structure as those of the first embodiment and the second embodiment are denoted by the same reference numerals, and descriptions thereof are omitted.

Based on the same atmospheric data 82c at the same point in time and the operating state information 82d obtained from the same device 82 as the atmospheric analysis environment sensor device 12 to be corrected, a highly accurate analyzer 89 is operated. Then, the data 89a obtained from the high-precision analysis device 89 is used as the true value, and the measurement data 82c of the uncalibrated environment sensor device 82 paired with the analysis device and the information 82d from the operating state information , the model is learned, the measured data 12a and the operating state information 12b obtained from the environment sensor device 12 to be calibrated are introduced into the model as measured data, and the calibrated data 18 is output. Other structures are the same as those of the first embodiment.

Further, the correction method of the present embodiment may use the correction method described in the first embodiment, or may be realized by a storage medium storing the software described in the first embodiment. .

[Other embodiments]
In another embodiment, the drift of the environmental sensor measurement data over time may be further taken into account and further corrected.

Specifically, when performing the above correction measurement, starting from the time of the previous accurate calibration and installation, the measurement data of the environment sensor may be acquired multiple times by recording different time points. The operating time of the environment sensor may be recorded by simultaneously obtaining the time and calculating the time difference. Then, the data measured a plurality of times and the corresponding time data are associated, the associated data is processed by the model of the neuron network, and the correction characteristics considering the factors over time are obtained.

With such a correction characteristic, drift of the environment sensor over time can be further corrected, and measurement accuracy can be further improved.

11 Environment sensor system in the first embodiment 12 Uncorrected atmospheric analysis environment sensor device in the first to third embodiments 12a Measurement data of the uncorrected environment sensor device 12b Measured by an environment sensor element that ensures accuracy and does not require correction operating state information of the environment sensor device (abbreviated as "operating state information")
13 Radio base station 14 Internet 15 Server device for data storage and correction processing 16 Database of environmental sensor measurement data published on the network 16a Same type as uncorrected environmental sensor device measurement data published on the network Measured data 16b Operating state information when measuring data published on the network 17 Learning data set used for learning model 18 Model for correcting measured data 18a Measured data after correction 21 Implementation model 22 For first correction 22an Output data obtained by correcting n-th uncorrected environmental sensor measurement data by a one-layer neural network 23 Neural network for second correction 23an n-th corrected measured data 24an n-th environmental sensor 1st layer neural network W1, W2 model matrix for initial correction of output data of element or module 24b Network model for 2nd layer neural network 23 31 1st neural network (24nn)
31a neural network input layer 31b neural network hidden layer 41 second neural network (24b)
41a Neural network input layer 41b Neural network hidden layer 41c Neural network output layer 52 Example of detected data 53 Example of processed data 61 Correlation between each measurement data before and after correction measured by the environment sensor device and the true value 71 Second Environment sensor system in the embodiment 72 Air analysis environment sensor device having the same configuration as the pre-corrected air analysis environment sensor device 72a Data measured by the pre-corrected environment sensor device 81 Environment sensor in the third embodiment System 82 Uncorrected atmospheric analysis environment sensor device of the same type as the atmospheric analysis environment sensor device paired with the analysis device 89 82c Measurement data of the same type as the data of the air analysis environment sensor device paired with the analysis device 82d Analysis device Operation status information of the same type as the data of the atmospheric analysis environment sensor device to be paired 89 Analytical device capable of acquiring the same type of data as the environment sensor ensuring high accuracy 89a Data measured by the analytical device

Claims (11)

A device for processing detection data of an environment sensor for processing measurement data obtained by an environment sensor device measuring the concentration of a detection target in the environment,
a measurement data input unit that acquires a plurality of learning measurement data of the environment sensor device in a learning mode and acquires actual measurement data of the environment sensor device in an actual measurement mode;
a true value acquiring unit that acquires true values respectively corresponding to the learning measurement data;
A plurality of association data are obtained by associating the plurality of learning measurement data with the corresponding true values, and a portion of the association data is randomly selected from the plurality of association data to create an association data set for learning. and processing the association data set using a neuron network model to obtain correction characteristics repeatedly a plurality of times, obtaining a plurality of correction characteristics, and obtaining a correction characteristic for each of the correction characteristics. a correction characteristic generating unit that verifies each correction characteristic based on a true value in association data that is not used in execution, and selects the optimum correction characteristic based on the verification result;
and a measurement data correction unit that corrects the actual measurement data based on the optimum correction characteristic. further comprising an operating state information input unit for inputting operating state information of the environment sensor device;
The operating state information includes at least one of internal and/or external temperature, humidity, atmospheric pressure, illumination level, ID information of the environment sensor device, and geographic location information of the environment sensor device when the environment sensor device operates. contains one piece of information,
2. The processing device according to claim 1, wherein when acquiring said association data, it further adds operating state information for association. The detection target is a PM value and/or a gas component,
2. The processing apparatus of claim 1, wherein the gas component includes at least one of nitrogen dioxide, sulfur dioxide, carbon monoxide, ozone, carbon dioxide, hydrogen sulfide, and VOCs. The processing device uses a neuron network model based on measurement data obtained by multiple types of sensor devices for different detection targets to add mutual interference characteristics between different detection targets to the correction features, and/or The plurality of learning measurement data and actual measurement data acquired from the environment sensor device by the measurement data input unit are learning measurement data and actual measurement data corrected in advance based on the operating state information by the sensor device, Alternatively, after the measurement data input unit acquires a plurality of learning measurement data and actual measurement data from the environment sensor device, the data pre-correction unit pre-corrects the learning measurement data and the actual measurement data based on the operating state information. 3. The processing apparatus according to claim 2, correcting and supplying to said correction characteristic generator. 2. The method according to claim 1, wherein the true value is at least one of data acquired by a verified environment sensor device, true data stored on the Internet, and data detected by a high-precision analysis device. Processing equipment as described. When the measurement data input unit and the true value acquisition unit acquire corresponding measurement data, the operation state information input unit further acquires time point data of the measurement data,
3. The processing device according to claim 2, wherein when acquiring said association data, said point-in-time data is further added to operation state information for association. further comprising an operating state information input unit for inputting operating state information of the environment sensor device;
The operating state information includes at least one of internal and/or external temperature, humidity, atmospheric pressure, illumination level, ID information of the environment sensor device, and geographic location information of the environment sensor device when the environment sensor device operates. contains one piece of information,
Before the correction characteristic generation unit acquires the plurality of correction features, the environment sensor device is corrected in advance based on the operating state information. Specifically, for each of the environment sensor devices, associated with the plurality of learning measurement data and the corresponding operating state information and the corresponding true value, and correcting the environment sensor device in advance based on the data obtained by the association;
2. The processing apparatus according to claim 1, wherein the correction characteristic generation unit uses the obtained pre-corrected data as the learning measurement data when acquiring the plurality of correction features. a plurality of sensor devices connected to a network;
a processing device according to any one of claims 1 to 7 connected to the network,
The environment sensor system, wherein the measurement data input unit of the processing device acquires actual measurement data of the plurality of sensors via a network and corrects the actual measurement data. The sensor device is a sensor device in which some or all of the sensor elements are not corrected, and/or the sensor device includes sensors that measure concentrations of a plurality of types of detection targets, respectively, and an operating state of the environment sensor device. 9. The environment sensor system according to claim 8, further comprising an operating state information acquisition unit group that acquires information with high precision. A method for processing detection data of an environment sensor for processing measurement data obtained by an environment sensor device measuring the concentration of a detection target in the environment,
a measurement data input step of acquiring a plurality of learning measurement data of the environment sensor device in a learning mode and acquiring actual measurement data of the environment sensor device in an actual measurement mode;
a true value acquisition step of acquiring true values respectively corresponding to the learning measurement data;
A plurality of association data are obtained by associating the plurality of learning measurement data with the corresponding true values, and a portion of the association data is randomly selected from the plurality of association data to create an association data set for learning. and processing the association data set using a neuron network model to obtain correction characteristics repeatedly a plurality of times, obtaining a plurality of correction characteristics, and obtaining a correction characteristic for each of the correction characteristics. a correction characteristic generating step of verifying each correction characteristic based on true values in association data not used in execution and selecting the optimum correction characteristic based on the verification result;
and a measurement data correction step of correcting the actual measurement data based on the optimum correction characteristics. A computer-readable storage medium storing a computer program,
When a processor included in the environmental sensor system executes the computer program, the environmental sensor system:
a measurement data input step of acquiring a plurality of learning measurement data of the environment sensor device in a learning mode, and acquiring actual measurement data of the environment sensor device in an actual measurement mode;
a true value acquisition step of acquiring true values respectively corresponding to the learning measurement data;
A plurality of association data are obtained by associating the plurality of learning measurement data with the corresponding true values, and a portion of the association data is randomly selected from the plurality of association data to create an association data set for learning. and processing the association data set using a neuron network model to obtain correction characteristics repeatedly a plurality of times, obtaining a plurality of correction characteristics, and obtaining a correction characteristic for each of the correction characteristics. a correction characteristic generating step of verifying each correction characteristic based on true values in association data not used in execution and selecting the optimum correction characteristic based on the verification result;
and a measurement data correcting step of correcting the measured data based on the optimum correction characteristics.

Images