Real-time ai-based calibration cloud system of 6-axis force / torque sensor

Abstract

The present invention relates to a calibration model construction method, a calibration model, and a 6-axis control device including a calibration model and provides a real-time AI-based calibration cloud system of a 6-axis force / torque sensor for reducing cross talk and hysteresis from the 6-axis control device, the system including: a detection module for detecting detection information generated by a detection unit in the 6-axis control device; and an AI model for processing the detection information received from the detection module by using a predetermined classification and calculation method.

Description

A real-time AI-based compensation cloud system for 6-axis force / torque sensors.

[0001] The present invention relates to a real-time AI-based correction cloud system for a 6-axis force / torque sensor.

[0002] Calibration is typically applied to various devices to compensate for errors in measurements and other operations. In particular, axes that move in multiple directions can be affected by external forces or mechanical errors, resulting in discrepancies between the measured values ​​and the output values ​​even though they correspond to the input values. This discrepancy increases as the number of axes increases. For example, in a device that drives six axes, which is a combination of linear and rotational displacements, any external force or self-weight applied in one direction will induce displacement in the other connected directions, resulting in errors in various directions.

[0003] These errors reduce the precision and durability of the device, making it difficult to use effectively in industrial environments. While various methods have been implemented to compensate for these errors, they do not occur within a consistent error range across various directions, but rather through complex error-causing factors. Therefore, proactive measures are needed to compensate for these errors.

[0004] One embodiment of the present invention aims to solve the problem of computational load of local equipment by cloudifying the increased computational amount for solution through an artificial intelligence correction algorithm.

[0005] One embodiment of the present invention aims to enable an artificial intelligence neural network security design with improved security by preventing neural network algorithms and functions from being stored in local storage media.

[0006] One embodiment of the present invention aims to implement simplification of local equipment without requiring design requirements for increased computational volume.

[0007] One embodiment of the present invention aims to enable linkage between a cloud server and a database.

[0008] One embodiment of the present invention aims to enable real-time data utilization through a database, thereby enabling real-time data information utilization regardless of physical distance.

[0009] The present invention relates to a calibration model construction method, a calibration model, and a 6-axis control device including the calibration model, and provides a real-time AI-based calibration cloud system for a 6-axis force / torque sensor for reducing crosstalk and hysteresis from a 6-axis control device, the real-time AI-based calibration cloud system for a 6-axis force / torque sensor including a detection module for detecting detection information generated in a detection unit in the 6-axis control device; and an AI model for processing the detection information received from the detection module using a predetermined classification and calculation method.

[0010] Additionally, the AI ​​model can be a model that processes dependent variables of at least two dimensions.

[0011] Additionally, AI models can process sensor information through DNN Regression (Deep Neural Network Regression).

[0012] In addition, the detection unit includes a first axis sensor, a first torque sensor, a second axis sensor, a second torque sensor, a third axis sensor, and a third torque sensor, and the detection information may include each detected information.

[0013] According to one embodiment of the present invention, a calibration model construction method for solving the problem of computational load of local equipment by cloudifying the increased computational amount for solution through an artificial intelligence correction algorithm, a calibration model, and a 6-axis control device including the calibration model can be provided.

[0014] According to one embodiment of the present invention, a calibration model construction method capable of designing an artificial intelligence neural network security with improved security by preventing neural network algorithms and functions from being stored in a local storage medium, a calibration model, and a 6-axis control device including the calibration model can be provided.

[0015] According to one embodiment of the present invention, a calibration model construction method that implements simplification of local equipment without design requirements for increased computational amount, a calibration model, and a 6-axis control device including the calibration model can be provided.

[0016] According to one embodiment of the present invention, a calibration model construction method capable of linking a cloud server and a database, a calibration model, and a 6-axis control device including the calibration model can be provided.

[0017] According to one embodiment of the present invention, a calibration model construction method capable of utilizing real-time data information regardless of physical distance by utilizing real-time data through a database, a calibration model, and a 6-axis control device including the calibration model can be provided.

[0018] Figures 1 to 5 illustrate a method for constructing a calibration model according to one embodiment of the present invention.

[0019] FIG. 6 is a conceptual diagram showing a detection module, a server, and a database as a configuration for performing a real-time AI-based correction cloud system of a 6-axis force / torque sensor according to one embodiment of the present invention.

[0020] Figure 7 is a conceptual diagram illustrating the application of an AI correction algorithm according to one embodiment of the present invention.

[0021] FIG. 8 is a graph showing force data by Linear Regression and a diagram showing force data by DNN Regression according to one embodiment of the present invention.

[0022] FIG. 9 is a diagram showing a graph showing hysteresis using linear regression according to one embodiment of the present invention and a graph showing hysteresis using DNN regression.

[0023] Hereinafter, specific embodiments of the present invention will be described with reference to the drawings. However, these are merely examples and the present invention is not limited thereto.

[0024] In describing the present invention, detailed descriptions of known technologies related to the present invention will be omitted if they are deemed to unnecessarily obscure the gist of the invention. Furthermore, the terms described below are defined based on their functions within the present invention and may vary depending on the intent or custom of the user or operator. Therefore, their definitions should be based on the overall content of this specification.

[0025] The technical idea of ​​the present invention is determined by the claims, and the following examples are merely a means of efficiently explaining the technical idea of ​​the present invention to a person having ordinary skill in the technical field to which the present invention belongs.

[0026]

[0027] Hereinafter, before explaining the calibration automation method of the present invention, a method for constructing a calibration model to be automated will be described through FIGS. 1 to 5. A calibration model that compares an output value arithmetically output from an input value input to each axis in a device controlling multi-axis calibration with an actual measurement value and compensates so that the measurement value converges to the output value will be described below. The calibration model can be applied to a multi-axis control device (400) to enable operation with improved precision.

[0028]

[0029] FIG. 1 is a flowchart illustrating a method for constructing a calibration model according to one embodiment of the present invention.

[0030] Referring to FIG. 1, a calibration model construction method according to an embodiment of the present invention includes a first module connection step (S10) in which a detection unit (100) connected to each axis included in a multi-axis is connected to a first module (200) that receives detection information detected from the detection unit (100), a data reception step (S20) in which the first module (200) receives detection information from the detection unit (100), a correction model construction step (S30) in which the detection information converted into a digital signal from the first module (200) is received by the learning unit (300) to construct a correction model, a first module application step (S40) in which the correction model constructed from the learning unit (300) is applied to the first module (200), a data storage step (S50) in which the correction model is stored in the first memory (222) of the first module (200), and the correction model applied to the first module (200) is stored in the data storage step (S50). It includes a correction data output step (S60) that processes the detection information transmitted from the detection unit (100) in a predetermined manner and outputs correction data.

[0031] The above steps are performed sequentially and include a process of constructing a correction model. The correction model can be constructed based on the detection information transmitted from the detection unit (100). Specifically, the multi-axes are axes in the X, Y, and Z directions in three dimensions, and the detection unit (100) may include a first axis sensor (111) that detects a force generated in an axial direction in the X direction, a first torque sensor (112) that detects a torque generated in an axis in the X direction, a second axis sensor (121) that detects a force generated in an axial direction in the Y direction, a second torque sensor (122) that detects a torque generated in an axis in the Y direction, a third axis sensor (131) that detects a force generated in an axial direction in the Z direction, and a third torque sensor (132) that detects a torque generated in an axis in the Z direction.

[0032] When a force is applied in one or more directions among multiple axes, the axes that are directly or indirectly linked to each other can transmit the force to the adjacent axes. Therefore, even if a force is transmitted in the X direction, for example, the force can also be influenced in the Y and Z directions. For example, displacement of one end of a control device (400) connected through multiple axes can occur in the Z and Y directions even when only a force is applied in the X direction. Furthermore, not only displacement in the axial direction but also torque occurring in the torsional direction of the axes can occur depending on the structure in which the multiple axes are connected, and the displacement of one end of the control device (400) can be changed by the torque.

[0033] If an abnormality in the first module (200) is detected in each of the aforementioned steps, the second module (500) may be driven to generate a correction model, and correction data may be generated using the generated correction module. The generated correction data is transmitted to the control device (400) so that the control device (400) can implement an output value corresponding to the input value.

[0034] The configuration for implementing the above calibration model construction method is as described with reference to Figure 2 below.

[0035]

[0036] Figure 2 is a conceptual diagram illustrating a calibration model according to one embodiment of the present invention.

[0037] Referring to FIG. 2, a calibration model according to one embodiment of the present invention includes a sensing unit (100), a first module (200), and a learning unit (300). Specifically, the sensing unit (100) includes an axis sensor that detects a force acting in at least one axial direction and a torque sensor that detects a torque acting on at least one axis. In addition, the first module (200) includes a first converter (210) that converts the sensing information detected by the sensing unit (100) from an analog state to a digital state, and a communication unit (221) that transmits the converted sensing information. In addition, the learning unit (300) includes a training unit (310) that repeatedly learns an error by comparing the sensing information transmitted from the communication unit (221) with an input value, and a transmission unit (320) that transmits a correction model constructed by being learned through the training unit (310) to the first module (200).

[0038] Here, the first module (200) may further include a first memory (222) that stores a correction model transmitted from the transmission unit (320). Accordingly, the first module (200) stores a correction model generated based on the detection information, and at the same time, receives the detection information transmitted from the detection unit (100) to the first module (200) from the first converter (210) and transmits correction data calculated through the correction model to a control device (400) including multiple axes. The correction data may be calculated through the correction model stored in the first memory (222) and transmitted to the control device (400).

[0039] The second module (500) described above can be operated in place of the first module (200) when the first module (200) is detected as being in an abnormal state. That is, the first module (200) and the second module (500) can be selectively operated. To this end, the second module (500) can be further included from the above-described embodiment, and the transmission path of the correction information model and the detection information can be changed. In this regard, this will be described in detail later with reference to FIGS. 3 and 4 below.

[0040]

[0041] FIG. 3 is a schematic diagram showing that data corrected through a constructed calibration model according to one embodiment of the present invention is transmitted to a control device (400), and FIG. 4 is a diagram showing a process of constructing a calibration model through a second module (500) replacing a first module (200) according to one embodiment of the present invention.

[0042] Referring to FIGS. 3 and 4, if the normal state of the first module (200) is detected after the correction data output step described through FIG. 1, the correction data is transmitted to a control device (400) including a multi-axis, and if an abnormal state is detected, a second module (500) that produces correction data by replacing the first module (200) may be further included. Here, the normal state may be determined based on satisfaction or dissatisfaction based on a predetermined error range.

[0043] The above-determined error range can be determined as an abnormal state when the error rate exceeds 0% or 10% in which the difference between the arithmetic output value according to the input value and the actual measurement value measured by the detection unit (100) is an error rate. Here, 0% is an ideal case where the output value and the measurement value are the same as the input value, but since the possibility is low, re-measurement can be performed through the second module (500) in the sense that re-measurement is requested. In addition, since a case exceeding 10% can be an upper limit rate of the error range selectively set by the user, it is not limited thereto.

[0044] The above measurement values ​​may be for comparison between the measurement values ​​detected by the first axis sensor (111) to the third axis sensor (131) and the first torque sensor (112) to the third torque sensor (132) and the output values. The first module (200) includes a first memory (222) in which a correction model transmitted from the learning unit (300) is stored, the detection unit (100) includes a built-in memory (140) in which a correction model transmitted from the first memory (222) is stored, and the second module (500) may further include a second memory (520) in which a correction model transmitted from the built-in memory (140) is stored.

[0045] In addition, the second module (500) further includes a second converter (510) that converts the detection information from an analog state to a digital state, and when the second module (500) is activated, the detection information from the detection unit (100) is transmitted to the second converter (510), and the converted detection information from the second converter (510) can transmit correction data to the control device (400) through the second memory (520).

[0046] Meanwhile, when the input value, which is output data, is performed through the control device (400), the correction model can continuously generate correction data from the detection information transmitted from the detection unit (100) and transmit it to the control device (400). That is, the flow for operating the control device (400) and the flow for generating the correction model may be different. In order to generate the correction model, there may be an operation flow that passes through the learning unit (300), and the flow for operating the control device (400) may have a flow that does not include the learning unit (300).

[0047]

[0048] And, when the second module (500) replaces the first module (200), as illustrated in FIG. 4, a correction model can be newly generated separately from the one generated through the first module (200). That is, the steps of connecting the second module (P10), receiving data from the first memory (222) (P20), constructing a correction model (P30), applying the correction model to the second module (P40), and outputting correction data (P50) can be sequentially performed. Outputting the correction data (P50) may be a process of transmitting output data to the control device (400). Although the second module (500) partially plays the role of the learning unit (300), it is separately distinguished, and the model generated in the learning unit (300) can be used as a correction model in the first module (200) or the second module (500).

[0049] This will be explained in more detail with reference to Figure 5 below.

[0050]

[0051] FIG. 5 is a schematic diagram showing the construction of a calibration model through a second module (500) according to one embodiment of the present invention.

[0052] Referring to FIG. 5, if the normal state of the first module (200) is detected after the correction data output step as described above, the correction data is transmitted to the control device (400) including the multi-axis, and if an abnormal state is detected, the correction data can be transmitted to the control device (400) after the execution step of the second module (500) in which the second module (500) that produces the correction data is performed.

[0053] The first module (200) includes a first memory (222) in which a correction model transmitted from the learning unit (300) is stored, the detection unit (100) includes a built-in memory (140) in which the correction model transmitted from the first memory (222) is stored, and the second module (500) may further include a second memory (520) in which the correction model transmitted from the built-in memory (140) is stored. That is, the correction model generated in the learning unit (300) can be shared by the first memory (222) and the second memory (520) to produce output data, and when the correction model is generated for the sharing, it can be configured to be backed up to the built-in memory (140), which is a component of the detection unit (100), through the first module (200).

[0054] According to this configuration, if a problem occurs in the first module (200) and normal operation becomes difficult, the correction module can be driven through the second module (500). At this time, the detection information from the detection unit (100) transmitted to the second module (500) is directly transmitted to the second converter (510) of the second module (500). That is, the detection information detected by the detection unit (100) can be transmitted from the second memory (520) of the second module (500) that stores the correction model, which is the backed-up information, to produce output data.

[0055]

[0056] The calibration model described above through FIGS. 1 to 5 is merely an example for explaining the automated calibration method of the present invention, and is not limited to various calibration models to which the automated method described below from FIG. 6 can be applied.

[0057] This technology applies an AI model (750) to compensate for the crosstalk phenomenon between sensors measuring force and torque and to obtain accurate final measurement values, and the following key technical elements may be required.

[0058] First, data collection and transmission technology involves collecting measured values ​​from each sensor and transmitting them reliably. Since the raw data obtained from each sensor can vary, it must be collected quickly and transmitted to subsequent processing stages. Furthermore, data alignment technology can align (synchronize) data collected from different sensors by compensating for temporal and spatial differences and assigning accurate timestamps.

[0059] Furthermore, AI correction algorithm technology can apply an AI correction algorithm based on sorted data. The algorithm detects crosstalk as a result of the influence between data detected and collected by each sensor, analyzes the influence of each sensor, and generates a corrected measurement value. In this case, it is preferable for the AI ​​model (750) to perform optimal correction by considering at least one of the interactions between sensors and data characteristics.

[0060] Furthermore, cloud-based computing technology is characterized by performing the necessary calculations for calibration in the cloud rather than on local equipment, enabling the execution of complex AI algorithms and processing of large amounts of data by leveraging high computing power in a cloud environment. The transmission of calibrated data is then transmitted back to the user or another system (such as a database (800)). This data can compensate for crosstalk and provide accurate force and torque measurements. Crosstalk refers to the phenomenon in which different signal channels transmit signals to surrounding channels, generating noise. Each input tool must be calibrated to set independent reference values ​​and parameters to match the intended use of the device to which the force sensor will be attached.

[0061] Specifically, calibration must be performed by independently measuring the three-dimensional axes, but since applying force to each axis also affects the other axes, a method for eliminating crosstalk occurring in the axes to be measured is required, and basically, the sensor has weights within the measurable range, and an AI model (750) that can compensate for hysteresis (a phenomenon in which a physical quantity is not restored to a previous value when a change occurs within a specific range) and crosstalk occurring when a physical quantity increases or decreases can be performed through a cloud server (700). This can be usefully used in various applications requiring automation and accurate force and torque measurement, and can provide a high degree of reliability and accuracy through cloud-based computing.

[0062] The AI ​​model (750) applied in the present invention is trained by outputting dependent variables of two or more dimensions, and updates the weight through Gradient Decent, which is a method of differentiating the loss function, but DNN regression can be applied to optimize the weight based on the average value calculated from the error of the entire learning data for each update.

[0063]

[0064] FIG. 6 is a conceptual diagram showing a detection module (600), a server (700), and a database (800) as a configuration for performing a real-time AI-based correction cloud system of a 6-axis force / torque sensor according to one embodiment of the present invention.

[0065] Referring to FIG. 6, the detection module (600) can transmit the detected detection information to the server (700). The information transmission between the server (700) and the detection module can be data transmission via TCP / IP. The detected detection information transmitted here can be information detected by the detection unit. That is, it can be information such as force detected from the first axis sensor, the first torque sensor, the second axis sensor, the second torque sensor, the third axis sensor, and the third torque sensor, and the detected information can be transmitted to the server (700) as is without separate processing. As described above, the server (700) includes an AI model (750), and the AI ​​model (750) can perform calculations on the detected information transmitted to the server (700).

[0066] Of course, a process of classifying data before performing an operation can be performed, and the classified data can be used to generate corrected data through a certain algorithm in the raw data state. This is an algorithm for correcting cross talk and hysteresis. Since it was confirmed that a difference from the true value occurs with the above-described Linear Regression (see Figures 8 and 9 below), it is desirable to apply a dependent model of two or more dimensions to correct this. For example, a regression method through MV Regression (Multivariable Regression) can be applied as shown in the figure below.

[0067] However, more preferably, by applying DNN Regression (Deep Neural Network Regression), a result almost identical to the true value can be obtained. That is, the conceptual diagram illustrating the application of the AI ​​correction algorithm according to an embodiment of the present invention through Fig. 7 below can correct raw data through the AI ​​model (750) included in the server (700). The results can be explained through the experimental data of Figs. 8 and 9 below.

[0068] Meanwhile, as illustrated in FIG. 6, the server can form a database (800), which is a collection of accumulated data through storage, by receiving the detected information, and can provide the corrected data received from the server (700) to the user through the detection module (600), thereby enabling it to be utilized in various ways.

[0069]

[0070] FIG. 8 is a graph showing force data by Linear Regression and a diagram showing force data by DNN Regression according to one embodiment of the present invention, and FIG. 9 is a graph showing hysteresis by applying Linear Regression and a diagram showing hysteresis by applying DNN Regression according to one embodiment of the present invention.

[0071] Meanwhile, the aforementioned DNN Regression shows visible effects as shown, but additional optimization can be performed through BSGD (Batch Stochastic Gradient Decent) or MBSGD (Mini Batch Stochastic Gradient Decent) depending on the user's choice.

[0072]

[0073] While representative embodiments of the present invention have been described in detail above, those skilled in the art will appreciate that various modifications to the above-described embodiments are possible without departing from the scope of the present invention. Therefore, the scope of the present invention should not be limited to the described embodiments, but should be defined not only by the claims set forth below but also by equivalents thereof.

Claims

In a real-time AI-based compensation cloud system for a 6-axis force / torque sensor to reduce crosstalk and hysteresis from a 1.6-axis control device, A detection module that detects detection information generated in a detection unit in a 6-axis control device; and A real-time AI-based correction cloud system for a 6-axis force / torque sensor, including an AI model that processes the detection information received from the detection module using a predetermined classification and calculation method.

2. In claim 1, The above AI model is a real-time AI-based correction cloud system for a 6-axis force / torque sensor, which is a model that processes dependent variables of at least two dimensions.

3. In claim 2, The above AI model is, A real-time AI-based correction cloud system for a 6-axis force / torque sensor that processes the above-mentioned detection information through the above-mentioned DNN Regression (Deep Neural Network Regression).

4. In claim 1, A real-time AI-based correction cloud system for a 6-axis force / torque sensor, wherein the detection unit includes a first axis sensor, a first torque sensor, a second axis sensor, a second torque sensor, a third axis sensor, and a third torque sensor, and wherein the detection information includes each detected information.

Images