Cloud ai digital level gauge

Abstract

To improve measurement accuracy of an inclination obtained by a digital level gauge using an acceleration sensor.SOLUTION: A digital level system includes a measuring instrument main body (11) including an acceleration sensor (21), an arithmetic device (22), and a first communication interface (23), the measuring instrument main body (11) being capable of acquiring raw data corresponding to an inclination with respect to a gravity direction using the acceleration sensor (21) and transmitting the raw data, a control device (12) configured to communicate with the first communication interface (23) and acquire the raw data, and a management server (13) configured to download a learned model obtained by learning a relationship between an inclination and raw data.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a high-precision level. [Background technology]

[0002] Digital levels equipped with sensors that can electronically acquire data on tilt are used to check the horizontality of various objects such as precision instruments and large-scale structures.

[0003] For example, Non-Patent Document 1 discloses a high-precision digital level that can output tilt values, which are digital data obtained with a resolution of 0.2 arcsec, to a device such as a personal computer. This level detects the direction of gravity using an acceleration sensor and outputs the detected value. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Akatsuki Manufacturing Co., Ltd., "DWL-9000XY - 2-Axis Ultra-Precision Digital Inclinometer / Level with Built-in Vibrometer | Akatsuki Manufacturing, the Top Level Manufacturer," [online], [Retrieved May 13, 2024], Internet<URL:https: / / www.kod-level.co.jp / products / digital-level2 / dwl-9000xy-digital-level / > Summary of the Invention [Problem to be solved by the invention]

[0005] The acceleration sensors installed in such digital levels can detect minute changes due to gravity. However, in order to output the data externally, these changes must be digitized according to a predetermined standard. However, the computing devices that can be installed in devices that can be used as digital levels have limited power supply and computing power, so the output values ​​must be truncated to a certain extent. As a result, the accuracy of the externally output values ​​cannot fully reflect the measurement accuracy that would be possible if the device's acceleration sensor were fully utilized.

[0006] SUMMARY OF THE INVENTION It is therefore an object of the present invention to further improve the measurement accuracy obtained by a digital level using a sensor such as an acceleration sensor. [Means for solving the problem]

[0007] As a first solution, this invention provides: a measuring device main body having an acceleration sensor, a computing device, and a first communication interface, the measuring device main body being capable of acquiring and transmitting raw data corresponding to an inclination with respect to a direction of gravity using the acceleration sensor; a control device that communicates with the first communication interface to acquire the raw data; a management server that enables downloading of a trained model obtained by training the relationship between the gradient and the raw data; The control device calculates a gradient corresponding to the raw data using the trained model downloaded from the management server. The digital level system solved the above problems.

[0008] Furthermore, in the first solution, the present invention provides: The relationship between the gradient and the raw data to obtain the trained model is A second solution can be adopted, which is the relationship between a group of raw data from the acceleration sensor measured with the measuring device main body facing in the direction of each tilt for a plurality of reference planes in different directions.

[0009] Furthermore, in the first or second solving means of the present invention, A third solution can be adopted in which the learning process includes basic learning to create a basic model for each design of the measuring device body, and a calibrated learned model in which the basic model is corrected for each individual measuring device body is used as the learned model used in the calculation.

[0010] The present invention also provides: a measuring device main body capable of acquiring and transmitting raw data corresponding to the tilt with respect to the direction of gravity using an acceleration sensor; a control device that acquires the raw data from the measuring device body, acquiring a trained model that has learned a combination of the tilt of the measuring device body and the raw data; Using the trained model, determining the slope corresponding to the raw data acquired by the measuring device main body; A fourth solution can be adopted as a tilt measurement method that performs the above. [Effects of the Invention]

[0011] The digital level of this invention learns the relationship between the raw data obtained by the acceleration sensor in the measuring instrument body and the actual corresponding tilt, and the control device can perform high-load processing using the learned model, making it possible to measure tilt with high accuracy without requiring excessive specifications in the measuring instrument body. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 is a functional block diagram illustrating an embodiment of a digital level system according to the present invention. [Figure 2] FIG. 1 is an overall flow diagram showing an example of a procedure for using the digital level system according to the present invention. [Figure 3] FIG. 1 is a partial flow chart showing an example of a procedure for basic learning of the digital level system according to the present invention. [Figure 4] Table showing an example of raw data sets uploaded from the control device [Figure 5] (a) Example of a pattern for creating a proofread trained model including elements, (b) Example of a pattern using the proofread trained model in (a) [Figure 6] (a) Example of a pattern for creating a calibrated trained model with elements removed, (b) Example of a pattern using the calibrated trained model in (a) DETAILED DESCRIPTION OF THE INVENTION

[0013] The present invention will be described in detail below with reference to embodiments. The present invention is a digital level system using an acceleration sensor and a method for measuring tilt.

[0014] The configuration of an embodiment of a digital level system according to the present invention will be described with reference to Fig. 1. This digital level system has a measuring instrument main body 11, a control device 12, and a management server 13.

[0015] The measuring device main body 11 is attached to a measuring object 14 whose tilt is to be measured, and is capable of acquiring data corresponding to the tilt. Specifically, it has an acceleration sensor 21, a calculation device 22, and a first communication interface 23 (first communication IF 23 in the figure).

[0016] The acceleration sensor 21 can detect acceleration corresponding to the direction of gravity, which changes depending on the tilt of the environment in which the measuring device main body 11 is installed. The detected results can be output as raw data in the three axial directions without being converted into angles, etc. Here, raw data refers to data in a state in which the acquired information remains, without truncating part of the data or lower digits or performing any conversion to reduce the number of digits.

[0017] The computing device 22 operates the acceleration sensor 21 and controls transmission of the acquired raw data via the first communication interface 23. There is no need to convert the acquired raw data on a large scale, and high computing power is not required. Rather, it is desirable for the computing device 22 to be able to operate with low power consumption.

[0018] The first communication interface 23 is a wired or wireless interface, and is an interface that allows mutual communication with the control device 12. Although a wired interface is possible, a wireless interface is preferable from the viewpoint of convenience. In particular, when a single control device 12 controls multiple measuring device bodies 11, a wireless interface can reduce the hassle of wiring. Among wireless interfaces, those that use low power are desirable, and Bluetooth (registered trademark) and Bluetooth Low Energy (Bluetooth LE) are preferably used. Bluetooth LE is particularly preferable from the viewpoint of power saving.

[0019] The measuring device main body 11 also has a power supply 24 for operation. If it is connected to an external AC power source, it may be subjected to vibrations via the cable, so it is desirable that it be a battery that can operate independently. As a battery, a lithium battery that achieves long life and high output can be preferably used, but is not particularly limited to this.

[0020] Furthermore, it is preferable that the measuring device main body 11 stores an ID that identifies each measuring device as information. If Bluetooth is used as the first communication interface 23, this ID may also serve as a Bluetooth ID. For example, a Bluetooth address, a unique ID, or a microcomputer serial number may be used. This ID is preferably attached to the raw data to be transmitted, and is preferably capable of being returned upon request.

[0021] The control device 12 is a device that can communicate with the measuring device main body 11, control the measuring device main body 11, and acquire raw data from the acceleration sensor 21. That is, the user controls the measuring device main body 11 via the control device 12 and obtains the tilt of the measurement target 14 as a measurement result from the measuring device main body 11. For this purpose, the control device 12 has a second communication interface 33 (second communication IF 33 in the figure) that can communicate with the first communication interface 23 of the measuring device main body 11. That is, the second communication interface 33 must be of the same standard as the first communication interface 23. The control device 12 also has a calculation unit 31 that performs various calculations for controlling the measuring device main body 11. A general CPU may be used as the calculation unit 31; that is, a general PC, smartphone, microcomputer, or the like may be used as the control device 12. The control device 12 also has both a temporary memory 32 used by the calculation unit 31 and a storage unit 36 ​​that stores data, programs, and the like. A magnetic disk, flash memory, or the like may be used as the storage unit 36. A trained model (described later) is stored in this storage unit 36 ​​to calculate measurement values. Furthermore, the control device 12 has an input device 34 for user operation and an output device 35 for the user to check the measurement results. The input device 34 can be a keyboard, mouse, trackball, touch panel, etc. The output device 35 can be a liquid crystal panel, organic EL panel, etc.

[0022] Furthermore, the control device 12 includes a network interface for connecting to a network. 37 (NWIF 37 in the figure). It is preferable to have a wired LAN terminal and a wireless LAN antenna for connecting to a local area network (LAN). In addition, it may be equipped with a communication device for connecting to a mobile communication network. In either form, it is preferable that it can communicate with the management server 13. Basically, it is preferable that it can be connected via the Internet, but it may also be connected via a dedicated closed network. The control device 12 uploads measured data and downloads trained models via the network interface 37.

[0023] Furthermore, the control device 12 is preferably installed with an application for controlling the measuring device main body 11 and for coordinating with the management server 13 as a program recorded in the storage unit 36, and is preferably capable of executing this application. By starting this application, the various operations of the digital level are performed.

[0024] The management server 13 is a server that can be connected to the control device 12 via a network. It does not have to be a single server, but may be a group of multiple servers, or may be a cloud server having the functions described below.

[0025] Management server 13 has an upload data storage unit 42 that can accept and store measurement data uploaded from control device 12. Upload data storage unit 42 may be stored integrally with management server 13, or may be stored in an accessible external storage unit in a different housing. The figure shows the data stored within management server 13. If this measurement data includes optional elements such as measurement conditions and the ID of measuring device main body 11, it will be easier to classify the conditions in the learning process described below.

[0026] The management server 13 registers trained models generated using measurement data and can download them in response to a request from the control device 12. It is preferable that multiple trained models with different measurement conditions and training conditions can be registered, and that a user of the control device 12 can select a suitable model from among them. For this selection and download, it is preferable that the management server 13 has a server function that can output a web page in cooperation with the above-mentioned app.

[0027] The management server 13 has a basic learning unit 43 that performs learning processing of a trained model using measurement data, either by itself or as a separate server connected to it. Because the learning processing requires high computing power, it is desirable that this unit be separate hardware from the server that responds to the control device 12.

[0028] Furthermore, the management server 13 has a download list unit 45 that stores the trained model generated by the basic training unit 43, or a correction model using the trained model, an additional trained model, etc. so that they can be downloaded from the control device 12. Specifically, it has a function as a web server connectable via the Internet, and preferably can respond to a request with a suitable trained model, etc.

[0029] The procedure for using data obtained by using the digital level system according to this embodiment will be described with reference to the overall flow in FIG. 2 and the detailed flow in FIG.

[0030] <Procedures for business operators> First (S100, Sa1), the measuring device main body 11 is designed, and a learning-only terminal that will serve as a reference model for the designed model number is fabricated (S101). This learning-only terminal will be used for future reference, so it is managed separately from the terminals to be shipped. There does not need to be just one learning-only terminal; it is preferable to use around 3 to 5 terminals as learning-only terminals in order to average out and absorb the individual characteristics of each terminal and to ensure there are spares.

[0031] For these learning-dedicated terminals, learning data to be used for basic learning is acquired (Sa2). Each terminal is used to orient itself at multiple angles in all three axial directions, and raw data output by the terminal is acquired based on the tilt of its orientation. Specifically, the angle (tilt) is placed on a reference plane that can be set with high precision (S102). Measurements are first taken on a horizontal plane (0°, 0°, 0°) (S110). The learning-dedicated terminal is then moved to 0°, 0.5°, 1°, etc. along the X-axis, 0°, 0.5°, 1°, etc. along the Y-axis, and 0°, 0.5°, 1°, etc. along the Z-axis (setting of the reference plane). Raw data obtained by acceleration sensor 21 in each state is acquired by control device 12 (S111-S132). The range of change shown here is an example, and may be smaller or with larger intervals. However, since too large an interval makes it difficult to ensure sufficient measurement accuracy, it is preferable that the interval be 1° or less. The raw data obtained by the acceleration sensor 21 should not be a single point for each angle condition, but should be acquired over a time series of 100 or more consecutive points, preferably 1000 or more, to ensure sufficient data for learning. For example, even if the tilt remains the same, if the tilt of the measurement target 14 continues to fluctuate slightly due to weak vibrations, even a difference of 1 / 100 of a second can change the raw data. The frequency of vibrations that can affect measurement differs between a 10 Hz and a 1000 Hz sampling rate. It is preferable to select a sampling rate that minimizes vibration. A measurement time of approximately 30 to 120 seconds is desirable for this sampling rate. By ensuring a certain amount of time, white noise can be countered and measurement time at a low data rate that learns the effects of power consumption and remaining battery life can be accommodated. However, prolonged operation rapidly drains the battery of the measuring device main unit 11, which is undesirable for product life, so it is desirable not to operate the device for too long. Furthermore, a sampling rate that is too fast also rapidly drains the battery, so a measurement time of 1000 Hz or less is preferable, and 400 Hz or less is even more preferable. For example, if the sampling rate is 10 Hz to 100 Hz, it is preferable to have 1000 or more data points as a countermeasure against white noise.

[0032] That is, a set of raw data obtained from a single measurement, which takes several seconds to perform in a time series, includes approximately 100 to 1,000 combinations of values ​​in the three axes obtained by the acceleration sensor. An example of this raw data is shown in FIG. 4. It is also preferable that the raw data be accompanied by the ID (terminal ID) of the learning-dedicated terminal used for identification. The control device 12 that acquires this raw data then associates highly accurate X, Y, and Z angle values ​​(reference plane angles) of the learning-dedicated terminal (measuring device main body 11) with the data. Here, raw data measured by varying the angle from 0 to 360 degrees in 0.5-degree increments in each of the X, Y, and Z axes is shown. In addition, the control device 12 may attach data such as other measurement conditions to the raw data. In the learning process described below, a trained model based on these conditions can be generated. Other conditions include, but are not limited to, measurement pace, temperature, humidity, altitude, latitude, and remaining battery charge. Among these, temperature and remaining battery charge, which tend to vary even at the same installation location, are particularly useful for correcting raw data. Obtaining data both before the start of measurement and after the end of measurement is preferable for improving accuracy. When measurements are long, data may be acquired during measurement, but this is preferably done when the measuring device main body 11 has a relatively high processing capacity and can ensure a sufficient data transfer volume. In addition to creating a trained model based on the conditions, this data may also be used to build a correction learning model based on this data, or it may be applied to an existing correction algorithm. Note that, as long as the raw data information is not lost on the receiving side, the raw data may be compressed or encoded when transmitted.

[0033] When multiple learning-dedicated terminals are used, raw data is acquired under similar conditions at different angles (S151 → Yes). The acquired data will have different terminal IDs, and the raw data itself will have different values ​​according to the characteristics of each device. Furthermore, when measurements are taken under different measurement conditions, the series of measurements up to this point is repeated again with different conditions (S161 → Yes → S162). These raw data sets with different angles and conditions are collectively called a raw data group.

[0034] After the control device 12 has acquired the raw data groups for basic learning, it uploads the raw data groups to the management server 13. Uploading does not need to be done for each measurement, and it is also possible to upload a large number of raw data groups with different angles, different conditions, or a large number of raw data groups with a different learning-dedicated terminal all at once. In other words, for all terminals managed as learning-dedicated terminals that are measuring device main bodies 11 of the same model number, raw data is acquired at different angles, and a raw data group corresponding to the angle value is generated and uploaded.

[0035] The management server performs basic learning using this uploaded group of raw data (Sa3). In other words, a trained model (basic model) is created that represents the basic characteristics when using the measuring device main body 11 of that model number. This basic model makes it possible to determine the degree of tilt of the X, Y, and Z axes that corresponds to the acquired raw data (xxxxx, yyyyy, zzzzz) when measuring with the measuring device main body 11 of that model number. In order to generate this basic model with as much accuracy as possible, it is desirable to use a basic model that combines values ​​from multiple learning-only devices. This makes it possible to absorb the individual characteristics of the learning-only devices.

[0036] The trained model and the basic learning method are not particularly limited. They may be models using neural networks, or models consisting of branching equations or mathematical formulas. However, because the amount of information is large, a method of deep learning of neural networks is preferably used. The trained model may be subject to transfer learning or fine tuning. In addition, it may be possible to detect whether there is an error between the raw data and the correct label (reference plane), and then perform separate learning to correct the correct label. Methods that use calculations based on mathematical programming or machine learning methods that incorporate the laws of physics (PIML) may also be used. Furthermore, separate learning may be performed for corrections based on data such as temperature and remaining battery capacity.

[0037] Then, when the measuring device 11 of that model number is shipped, calibration learning is performed on all products by the calibration learning unit 44 (Sa4). Since each measuring device 11 has slightly different characteristics, even though it is based on the basic model, differences from the basic model are learned. For example, when the X-axis, Y-axis, and Z-axis are all at 0 degrees, it is advisable to correct the deviation from the basic learning model. Using the calibration-learned model obtained in this way, it is possible to connect the relationship between the raw data and tilt for each individual measuring device 11 and determine the measured tilt value.

[0038] Furthermore, it is preferable for the operator to use groups of raw data with different measurement conditions, perform calibration and learning under different conditions, and prepare trained models suited to the measurement conditions. The operator registers in the management server 13 a calibrated and trained model corresponding to a basic model tailored to each condition (Sa5), and the user can selectively download and use the calibrated and trained model tailored to their usage environment from the management server 13 via the control device 12. Here, measurement conditions include sampling rates with different frequency bands for detecting vibrations, latitude, altitude, etc. For example, in a measurement environment where weak vibrations of a specific frequency are present, when measurements are to be taken under conditions of a sampling rate that can ignore that frequency band, a trained model suited to the sampling rate to be measured can be selected. A higher sampling rate is not necessarily preferable, but conditions suited to the actual usage conditions of the user are preferable.

[0039] This "calibrated, trained model" is posted on the download list section 45. The download list section 45 is also an externally accessible web server. It may be part of the management server 13 as shown in the figure, or it may be a separately provided server. The control device 12 can access this download list section 45 and select from the list and download the calibrated, trained model that corresponds to the ID of the measuring device main body 11 used by the control device 12. The control device 12 uses the downloaded calibrated, trained model to process raw data from the measuring device main body 11 corresponding to the ID using the calibrated, trained model to obtain measurement values. However, the procedure for reflecting the remaining battery charge of the measuring device main body 11, the temperature at which it is placed, and other conditions in the measurement values ​​obtained by the calibrated, trained model differs depending on whether the operator or the user reflects the information.

[0040] In addition, when the trained model that has undergone basic training is upgraded, it is advisable to prepare a calibrated trained model that reflects the calibration learning for each measuring instrument main body 11 that matches the upgrade.

[0041] A list showing differences in elements such as version differences and machine learning conditions for each model is displayed in the download list section 45. For example, assume that proofreading-trained models with different basic learning conditions are registered as shown in the following list. · Ver.0.1.0 Proofread Model (Basic Learning A + Proofreading) · Ver.0.2.0 Proofread Model (Basic Learning B + Proofreading) The download list section 45 sends a web page that displays the following list on the browser or app of the control device 12 according to the ID of the measuring device main body 11 to which the control device 12 is connected, and makes it available for download from an appropriate link. (ver.0.1.0, ID) (for the measuring instrument 11 of the corresponding ID, calibration-trained model of ver.0.1) (ver.0.2.0, ID) (for the measuring instrument 11 of the corresponding ID, calibration-trained model of ver.0.2) Also, different models may be prepared depending on the difference in proofreading learning.

[0042] This is the procedure that the system operator and manufacturer should follow. A user who wishes to use the present invention will use this environment to measure the tilt using the measuring device main body 11 and the control device 12.

[0043] In the above flow, the steps from Sa4 to Sa5 can take a different form. When preparing a trained model suitable for each terminal through calibration learning, instead of performing 0-degree calibration, or in addition to 0-degree calibration, an additional trained model can be used in which a group of raw data from each measuring device main body 11 is additionally trained on a basic model. As a method for this additional learning, a method that enables learning with low power consumption by changing only part of the trained model rather than changing the entire model, such as transfer learning or fine tuning, is preferably used.

[0044] <User procedure> A user intending to measure the tilt of a measurement target 14 using the measurement device 11 installs the shipped measurement device 11 on the measurement target 14 whose tilt is to be measured (Sa6) and then controls the control device 12 to acquire raw data (Sa7). Around this time, a trained model conforming to the measurement conditions is downloaded from the management server 13 and stored in the control device 12 (Sa8). From the perspective of accuracy of the calculated tilt, it is desirable to select a trained model that corresponds to the individual ID of the measurement device 11 and that has been corrected for calibration learning. When the control device 12 acquires the raw data from the measurement device 11, it uses the trained model (calibration-trained model) to calculate the tilt angle corresponding to the raw data as a measured value (Sa9). This allows the angle of the measurement target 14 on which the measurement device 11 is installed to be determined with extremely high accuracy.

[0045] The procedure for using a proofread and trained model, which is a trained model based on proofreading, is to follow the procedure corresponding to the generation of the proofread and trained model. Specific examples are given below.

[0046] <Pattern 1: Single-use model> Figure 5(a) shows an example of steps to be taken during the operator's learning phase. First, raw data (unprocessed raw data) from the acceleration sensor, without noise removal, is acquired from the measuring device main unit 11 as learning data. This unprocessed raw data records the changes over time on the X, Y, and Z axes, as well as the decrease in battery power and temperature over time. Note that while the horizontal axis represents time, the X, Y, and Z axes of the acceleration sensor share a common time axis, but the measurement data rate differs from the time axes of the battery power and temperature. This is because the changes in battery power and temperature are less rapid than those of the acceleration sensor, and the number of data points required differs. Specifically, the measurement intervals for battery power and temperature can be adjusted from several seconds to 5 to 10 minutes. In particular, when the measurement data rate of the acceleration sensor is high, even if the raw data is transmitted as is using this invention, the measurement and Bluetooth communication alone can impose a heavy load. Furthermore, if the intervals between temperature and remaining battery capacity measurements are too short, there is a risk that the remaining battery capacity will be consumed more quickly, so it is desirable to measure the remaining battery capacity at intervals of at least 10 seconds. An interval of about 1 minute is more realistic. By appropriately performing interpolation and prediction processes, the measurement intervals for temperature and remaining battery capacity can be extended, thereby reducing the consumption of remaining battery capacity.

[0047] In this example, the raw data is transmitted from the operator's control device 12 to the management server 13, which then performs noise removal on the unprocessed raw data from the acceleration sensor and preprocesses it to predict the remaining battery charge and temperature. The noise removal can be performed using a separate noise removal mechanism or a predetermined algorithm. For example, since acceleration sensor measurements contain white noise and noise from fluctuations in the measurement environment, removing this noise using digital signal processing such as smoothing or a low-pass filter can make the data more usable.

[0048] Meanwhile, temperature prediction and remaining battery capacity are interpolated or predicted using separate algorithms. Temperature is basically predicted based on the displacement of the temperature sensor built into the measuring device main body 11. As extended settings, room temperature settings and air conditioning settings acquired manually from the control device 12 or input from a separately installed indoor temperature sensor are used, and the management server 13 makes predictions by incorporating these settings into the temperature sensor value of the measuring device main body 11. For example, in a factory, it is recommended to automatically acquire room temperature settings and air conditioning settings. When the temperature changes drastically, such as immediately after turning on the heater indoors in winter, the measuring device main body 11 itself may be cold, resulting in a large difference from the outside temperature and a large temperature change during measurement. Because such temperature changes affect the sensitivity of the acceleration sensor, appropriate standardization is desirable.

[0049] The remaining battery charge has less of an effect on the acceleration sensor values ​​than temperature. This is because, when a voltage regulator is used as the measuring device main body 11, the voltage is boosted at startup, so as long as there is enough battery charge remaining to stably boost the voltage, there is little risk of operation being affected. However, if the battery charge is insufficient, stable operation of the acceleration sensor becomes difficult. Therefore, it is desirable to estimate the measurement time and measurement data rate that will enable stable measurement, taking into account the measurement load in addition to the remaining battery charge and boost, and then perform measurements within that stable operating range.

[0050] The preprocessed raw data is used to fine-tune the basic trained model. In this pattern, related factors such as offset and sensitivity are learned together, along with the nonlinearity of the remaining battery charge and temperature, to create a calibrated trained model. The data is a multidimensional space containing these learning factors. The calibrated trained model created in this way is posted in the download list section 45 and made available for download.

[0051] An example of the steps in the user's usage phase is shown in Figure 5(b). The calibration-trained model created using Pattern 1 above is a model that includes the remaining battery charge and temperature changes of the measuring device 11, as well as other learned elements. Therefore, the unprocessed raw data for measurement acquired by the user's control device 12 from the measuring device 11 is simply preprocessed to remove noise using the same procedure as used to create the calibration-trained model, and is converted into preprocessed raw data, eliminating the need for correction based on conditions such as the remaining battery charge and temperature. By using the calibration-trained model directly on the preprocessed raw data, it is possible to obtain measured values ​​that are optimized for the measuring device 11.

[0052] The calibrated trained model in Pattern 1 requires comprehensive training for the model to be used directly. This requires a large number of different trained models to be prepared, placing a heavy burden on the business. On the other hand, users can simply download a trained model that matches the measurement situation and use it as is, enabling use with guaranteed accuracy.

[0053] <Pattern 2: Pre-processing post-use model type> Figure 6(a) shows an example of the steps in the learning stage on the business side for a pattern in which the various elements described above are reflected on the user side. First, as in Pattern 1, raw data (unprocessed raw data) that has not been subjected to noise removal is acquired from the measuring device main body 11 as the data to be used for learning. This unprocessed raw data records the changes in the X-axis, Y-axis, and Z-axis over time, as well as the decrease in battery remaining capacity and temperature changes over time. The management server 13, which receives the raw data from the business-side control device 12, performs preprocessing on this unprocessed raw data, removing noise and predicting the changes in battery remaining capacity and temperature. For noise removal, a separate learning model for noise removal may be created and processed using artificial intelligence, or a predetermined algorithm may be used. Up to this point, the process is the same as in Pattern 1.

[0054] Next, the management server 13 removes (1) the nonlinear element due to the remaining battery charge from the preprocessed raw data. Then, (2) the nonlinear element due to the temperature is removed. Note that these are merely examples, and when creating a calibrated, trained model that excludes other nonlinear elements, they are similarly corrected and removed at this stage. The method for removing these nonlinear elements may involve establishing and using a correction formula according to the characteristics of the measuring device main body 11, or it may involve creating and using a separate machine learning model for excluding these elements.

[0055] Next, using the raw data from which the influence of these elements has been removed, the nonlinear relationship between the remaining element, the offset, and the sensitivity of each axis is learned to create a calibrated trained model. Unlike the calibrated trained model in Pattern 1, there is no need to prepare models that account for differences in remaining battery charge or temperature. Therefore, the number of calibrated trained models to be prepared is smaller than in Pattern 1, reducing the burden on the operator. As with Pattern 1 above, this calibrated trained model is posted on the download list unit 45 and made available for download.

[0056] An example of the steps in the user-side usage stage is shown in Figure 6(b). As with pattern 1, the user-side control device 12 also accesses the download list unit 45 and downloads the calibrated and learned model. When performing measurements with the control device 12, the raw data obtained from the measuring device main body 11 is not directly subjected to the calibrated and learned model, but rather the above-mentioned (1) nonlinear elements due to the remaining battery charge and (2) nonlinear elements due to temperature are removed to obtain element-removed raw data, and this element-removed raw data is then processed using the calibrated and learned model to obtain measurement values ​​in a two-stage process.

[0057] The calibrated trained model in Pattern 2 is trained without including information on remaining battery charge and temperature changes, so the number of training patterns can be reduced compared to Pattern 1. This reduces the burden on the operator. However, since correction processing is included as preprocessing, accuracy is slightly lower compared to Pattern 1. [Explanation of symbols]

[0058] 11 Measuring instrument body 12 Control device 13 Management Server 14 Measurement Object 21 Acceleration sensor 22 Arithmetic unit 23 First communication interface 24 Power supply 31 Arithmetic section 32 Temporary Memory 33 Second communication interface 34 Input Devices 35 Output Device 36 Preservation Department 37 Network Interface 42 Upload data storage section 43 Basic Learning Department 44 Proofreading Learning Section 45 Download List Section

Claims

1. a measuring device main body having an acceleration sensor, a computing device, and a first communication interface, the measuring device main body being capable of acquiring and transmitting raw data corresponding to an inclination with respect to a direction of gravity using the acceleration sensor; a control device that communicates with the first communication interface to acquire the raw data; a management server that enables downloading of a trained model obtained by training the relationship between the gradient and the raw data; The control device calculates a gradient corresponding to the raw data using the trained model downloaded from the management server. Digital level system.

2. The relationship between the gradient and the raw data to obtain the trained model is The relationship between the raw data group of the acceleration sensor measured with the measuring device main body facing in the direction of each tilt for a plurality of reference planes with different orientations, The digital level system of claim 1 .

3. the learning process includes basic learning to create a basic model for each design of the measuring device main body, and a calibrated learned model obtained by correcting the basic model for each individual measuring device main body is used as the learned model to be used for the calculation.

3. The digital level system according to claim 1 or 2.

4. a measuring device main body capable of acquiring and transmitting raw data corresponding to the tilt relative to the direction of gravity using an acceleration sensor; a control device that acquires the raw data from the measuring device body, acquiring a trained model that has learned a combination of the tilt of the measuring device body and the raw data; Using the trained model, determining the slope corresponding to the raw data acquired by the measuring device main body; Perform the tilt measurement method.

Images