Method for inferring force from a sensor device, method for training multiple networks, force inference module, and sensor device.

Abstract

We will improve the configuration of sensors for robotics. [Solution] The present invention relates to a method for inferring force from a force-sensing sensor device. This method comprises reading out multiple pressure values ​​and calculating a force map using a feedforward neural network. The present invention further relates to multiple methods for training multiple neural networks, a force inference module, and a sensor device.

Description

Technical Field

[0001] The present invention relates to a method for inferring the force of a sensor device, a method for training a plurality of networks, a force inference module, and a sensor device.

Background Art

[0002] When developing applications such as robots, detecting the force applied to the hand, leg, or other part of the robot such as an operating device is important for enhancing the function of the robot to move around or manipulate objects. Known implementations of sensor devices that can be used in robotic applications to obtain feedback regarding the applied force are very expensive and do not have sufficient resolution.

[0003] Such a device would be used for force measurement. However, known sensor devices require a high density of sensors (arrangement) to achieve high spatial resolution.

Summary of the Invention

Problems to be Solved by the Invention

[0004] Therefore, the problem of the present invention is, unlike the prior art, to provide a method for inferring the force of a sensor device optimized with respect to the prior art and related methods. A further problem is to provide a force inference module that executes those methods. A further problem is to provide a sensor device that detects force and has such a force inference module.

Means for Solving the Problems

[0005] This problem is achieved by the subject matter of the main claim. Preferred embodiments can be derived, for example, from the dependent claims. The content of the claims becomes the content of the description by express reference.

[0006] The present invention relates to a method for inferring the force of a sensor device that detects force.

[0007] Such a sensor device, in particular a sensor device that can use this method, may include multiple barometric pressure sensors. It may further include a tracking layer, which may cover the barometric pressure sensors and provide a measuring surface. For example, such a sensor device that can use the method of the present invention may be a sensor device as described herein, or can be manufactured according to the method described herein. With respect to the sensor device or manufacturing method, all of the disclosed embodiments and modifications may be used.

[0008] The method for inferring force is, - A step of reading the pressure value from the barometric pressure sensor, - A step of calculating a force map on a measurement surface based on pressure values ​​using a feedforward neural network, wherein the force map comprises multiple force vectors. It is equipped with.

[0009] Using this method, force inference from a barometric pressure sensor can be performed to obtain at least one of high resolution and advanced information. This is possible because feedforward neural networks have been shown to provide force information with a resolution much finer than the spacing of the barometric pressure sensors. Further information can also be provided (by using this method). This capability is particularly obtainable when the feedforward neural network is properly trained. Preferred (multiple) implementations of training are further described below.

[0010] A barometric pressure sensor can be adapted to generate an output signal that depends (for example, linearly) on the pressure applied to it. In particular, since the pressure is relayed from the measurement surface to the barometric pressure sensor, and typically even a force applied over a minimal area of ​​the measurement surface is relayed to several barometric pressure sensors, high resolution can be obtained using techniques such as feedforward neural networks.

[0011] With regard to the sensor device, refer to the detailed description, including the descriptions of embodiments and modifications shown herein.

[0012] The force map may be a map defined on an actual measurement surface. Here, the force map may have multiple map points. At each map point, some information, such as force vectors as further described below, may be defined. The force map typically provides information about the forces applied to the measurement surface. For example, such forces may arise from an indenter or several indenters pressing against the measurement surface, or from an object currently being manipulated by a sensor device (for example, if the sensor device is the end effector of a robot).

[0013] In one implementation, the feedforward neural network comprises a transfer network and a reconstruction network. The transfer network maps the barometric pressure sensor to multiple virtual sensors in a finite element model of the sensor device. The reconstruction network maps the virtual sensors in the finite element model to a force map. Each virtual sensor may have one or more virtual sensor points, each with its own value.

[0014] Therefore, the feedforward neural network is split in this implementation. This improves its functionality, and in particular, the training potential, as will be explained further below.

[0015] The transfer network can map actual barometric pressure sensors, or output values ​​generated from barometric pressure sensors, to a finite element model. The finite element model may be a virtual model of the actual sensor device. This may be used to enhance the force inference capability. The finite element model can be modeled using the finite element method. It (the finite element model) may have virtual representations of the actual components and materials used. For example, the Young's modulus and Poisson's ratio of the materials used can be the same as those of the actual sensor device. Distances and other geometric dimensions may also be identical between the actual sensor device and the virtual finite element model. However, it should be noted that the finite element model is a component mainly used for training and does not necessarily need to be implemented in an implementation used only for force inference after training has been performed. Once training has been performed, the transfer network and the reconstruction network can be used separately from the complete finite element model. In this case, force inference should be performed in each case; that is, a force map should be obtained, where the output values ​​read from the barometric pressure sensor are first mapped to the values ​​of the virtual sensor points by the transfer network, and the values ​​of the virtual sensor points thus obtained are mapped to the force map by the reconstruction network.

[0016] Typically, transfer networks and reconstructed networks are artificial neural networks. Mapping may mean that input values ​​are fed into the network, and the network generates output values ​​based on training. Training may also involve adapting a set of values ​​that define the network's behavior. For example, approximately one million numbers may be used to define the network's behavior. In the case of a transfer network, it may be fed data from a barometric pressure sensor, and this transfer network may generate values ​​for virtual sensor points. In the case of a reconstructed network, it may be fed values ​​for virtual sensor points, and this reconstructed network may generate a force map. The entire feedforward neural network, whether partitioned or not, is fed data from a barometric pressure sensor, and this feedforward neural network generates a force map.

[0017] A virtual sensor can be considered a segment of an actual sensor to a sensor point. While an actual sensor, such as a barometric pressure sensor, converts a single force applied to it into a single output signal, a virtual sensor may convert such a force into values ​​at multiple sensor points. Typically, the sensor points are located in regions of a finite element model corresponding to the barometric pressure sensor in the actual sensor device. The region within the finite element model may be, for example, around 10% or 50%. The concept of virtual sensors also takes into account the fact that barometric pressure sensors are not always located at points where their position is known with sufficient precision for precise positioning to be used in force inference. Using a finite element model with virtual sensor points enables reliable force inference despite such variations.

[0018] The following section describes the aspects of network training. The training steps mentioned in this section should be considered as steps performed before the actual force measurement and force inference are carried out. Therefore, the force inference method can be considered a combination of training steps performed before force inference and force inference using the trained network. The force inference method can be considered as force inference itself, using one or more appropriately trained networks. Furthermore, individual training methods are described below. These can be performed independently of any force inference. Force inference, typically involving the reading of a pressure sensor and the generation of a force map, can be considered an action performed in a particular use case. A "certain use case" is when the sensor device is used to measure or evaluate the force applied to the measurement surface, because the sensor device is currently manipulating an object or otherwise in contact with an object that applies pressure to the measurement surface.

[0019] In one implementation, the reconstruction network can be trained by the following steps, which are performed before force inference: - A step of performing multiple simulations in a finite element model, each simulation including simultaneously applying one or more simulated forces to a simulated measurement surface of the finite element model, thereby calculating a simulated force map on the simulated measurement surface, the simulated force map comprising multiple simulated force vectors, and calculating the values ​​of corresponding virtual sensor points using the finite element model. - A step to train the reconstructed network using the calculated simulated force map and the corresponding values ​​of the calculated virtual sensor points. That is the case.

[0020] Such training steps for the reconstruction network can be used to properly train the reconstruction network, enabling it to generate accurate and detailed force maps. These force maps represent the intended output of the sensor device, derived from the values ​​of virtual sensor points. The values ​​of virtual sensor points can be obtained, in particular, by the transfer network.

[0021] It has been proven that using only simulations is appropriate for training reconstruction networks. In particular, such simulations can be used to train reconstruction networks to detect not only one force but multiple forces acting on a measurement surface. This is far simpler than training the network through actual force tests. In actual force tests, applying two or more forces simultaneously is complex due to collision avoidance problems and the complex experimental setup. It has been shown that using only simulations to train reconstruction networks provides high reliability in reconstructing force maps. The simulations may be performed on a computer, or on another programmable entity and / or automated data processing entity.

[0022] Training of a finite element model can be performed purely by computer simulation. Therefore, simulated forces are also applied only in such computer simulations. The simulated measurement surface is typically the surface of the finite element model (e.g., the flexible layer of the finite element model). Therefore, the simulated measurement surface also exists only in simulations. Here, the measurement surface is the surface of the actual sensor device.

[0023] Simulated forces are applied to a simulation on a simulated measurement surface. This creates a simulated force map. The simulated force map comprises multiple simulated force vectors, each of which gives a local value for the simulated force map. The simulated force map can be represented and / or computed as a deformation of the simulated measurement surface. In particular, it can be computed using the finite element method.

[0024] The values ​​of virtual sensor points can also be calculated using the finite element method. In particular, the simulated forces and structural and material properties of a finite element model representing an actual sensor device determine both the simulated force map and the values ​​of virtual sensor points. This allows us to obtain the relationship between the simulated force map and the values ​​of virtual sensor points.

[0025] In force inference, the values ​​of virtual sensor points are generated based on data from barometric pressure sensors that indirectly measure actual forces. The relationship between the virtual sensor point values ​​and the simulated force map obtained from the simulation allows the force map to be reconstructed by a reconstruction network.

[0026] Note that generating a force map from virtual sensor point values ​​is referred to as reconstruction. Therefore, a network that performs such reconstruction is called a reconstruction network.

[0027] When training a reconstructed network, data from a run simulation may be used. Such data may include, in particular, simulated force maps and corresponding virtual sensor point values.

[0028] In one implementation, the simulated force applied to the simulated measurement surface is generated based on each simulated indenter, which has a simulated indenter shape. The shape may be particularly relevant to the portion of the simulated indenter that contacts the simulated measurement surface in the simulation. Therefore, the simulated indenter is the object used in the simulation to define the simulated force.

[0029] In one implementation, the simulated indenter shape is selected from the group comprising at least a tip, a circle, a triangular cross-section, a square cross-section, a hemisphere, a cube, and a cylinder. Such simulated indenter shapes have been proven appropriate because they correspond to the typical shapes of the actual object that comes into contact with the measurement surface when applied. Using such different indenter shapes greatly improves the training of the reconstruction network to reconstruct the corresponding or similar shapes applied to the actual measurement surface. Note that you can use each of the shapes mentioned, use only one of the shapes mentioned, or use a selection of the shapes mentioned. Alternatively, or in addition to them, other shapes can be used. If multiple indenters are used in the simulation, the shapes of the multiple indenters may be identical or different.

[0030] In one implementation, the reconstruction network was trained using multiple different simulated indenter shapes. This allowed the reconstruction network to be trained to distinguish between the forces generated by different indenter shapes. In particular, one or more simulations could be performed for each indenter shape used, or for each combination of indenter shapes used. Such simulations may differ, for example, in at least one of the following: the number of indenters and the locations where one or more indenters are applied.

[0031] In one implementation, the reconstruction network was trained using multiple sizes of simulated indenters. Furthermore, instead of using different shapes, the reconstruction network can be trained to distinguish between indenters or other objects applying forces of different magnitudes. For example, contact portions of different sizes may be used on the simulated measurement surface. The descriptions of running the simulation given for the use of different indenter shapes apply accordingly. Also, various combinations of indenter shapes and sizes are possible.

[0032] In one implementation, the reconstruction network was trained by at least part of a simulation that involved the simultaneous application of simulated forces generated based on two or more simulated indenters. This allows the reconstruction network to be trained to distinguish between forces applied by a single indenter and forces applied by two or more indenters. This is particularly feasible in simulation and is far simpler than preparing an experimental setup to perform such applications of two or more indenters.

[0033] In one implementation, the reconstruction network was trained by at least part of a simulation that included the application of simulated forces generated based on only one simulated indenter. This allowed for specific training for reconstructing the force map when only one indenter was applied.

[0034] For example, in typical training, the following number of simulations can be performed:

[0035] When training a reconstructed network with a single contact, simulations can be run 10,000 to 50,000 times, or up to 30,000 times.

[0036] When training a reconstructed network with two contacts, it is possible to run 5,000 to 20,000 simulations, or 10,000 simulations.

[0037] When training a reconstructed network with three contacts, it is possible to run 5,000 to 20,000 simulations, or 10,000 simulations.

[0038] When training a reconstructed network with four contacts, it is possible to run 5,000 to 20,000 simulations, or 10,000 simulations.

[0039] When training a reconstructed network with 5 contacts, it is possible to run 5,000 to 20,000 simulations, or 10,000 simulations.

[0040] However, these are merely typical or recommended values. In general, any number of simulations can be performed. For example, two contacts mean the simultaneous application of two forces, three contacts mean the simultaneous application of three forces, four contacts mean the simultaneous application of four forces, and five contacts mean the simultaneous application of five forces. Such simulations can be combined during training.

[0041] In one implementation, each simulated force vector comprises a normal force component, a first shear force component, and a second shear force component. Therefore, the force map provides information about these components. It should be noted that in typical prior art implementations, shear forces were not reconstructed. However, it has been shown that when simulated force vectors with such components are used to train a simulation-based reconstruction network, shear forces can be reconstructed in addition to normal forces. This provides valuable additional information for multiple applications, such as robotics applications manipulating objects.

[0042] In this implementation, the first shear force component of the simulated force vector corresponds to the first shear force, and the second shear force component corresponds to the second shear force. In particular, the first shear force is perpendicular to the second shear force. This provides easily usable information based on the perpendicular direction of the shear force.

[0043] Please note that force vectors can alternatively have around three components.

[0044] In one implementation, the reconstruction network was trained using multiple simulated forces with different shear force components. This allows the reconstruction network to be trained to distinguish between various shear forces applied to the measurement surface. The shear force can vary between different components of the force used in one simulation or across multiple different simulations.

[0045] In one implementation, the reconstruction network was trained using multiple simulated forces with different normal force components. This allows the reconstruction network to be trained to distinguish between various normal forces applied to the measurement surface. The vertical force can vary among the various forces used in one simulation or in multiple different simulations.

[0046] It should be noted that the concept of using at least one of different simulated indenters, different simulated indenter shapes, different simulated indenter sizes, and different simulated shear forces or different simulated shear force components can also be applied in other circumstances and contexts when training a neural network for the purpose of force inference. This is independent of the implementation of the sensor device given herein. The same applies to at least one of the actual indenter and multiple forces.

[0047] In one implementation, the transfer network may be trained by the following steps, which are performed before force inference: - A step of performing multiple force tests on a sensor device, each force test including applying a force with one indenter at a position on the measurement surface of the sensor device, simultaneously measuring the force applied by the indenter, and simultaneously measuring a pressure value using a barometric pressure sensor, - For each force test, the step of performing a corresponding simulation using the finite element model, wherein each simulation includes applying a simulated force to the simulated measurement surface of the finite element model, thereby calculating a simulated force map on the simulated measurement surface, the simulated force map including multiple simulated force vectors, the simulated force corresponding to the measured force and the force applied to a position on the simulated measurement surface corresponding to a position on the measurement surface, The steps include: calculating the values ​​of the corresponding virtual sensor points using a finite element model; - A step of training a transfer network with the measured pressure value and the corresponding calculated virtual sensor point value. That is the case.

[0048] Force testing, in contrast to simulation, is a test performed with actual physical sensor equipment. The indenter may be an object specifically designed to contact the measurement surface. Force testing can be performed with the sensor equipment moving relative to a stationary indenter, or with the indenter moving relative to a stationary sensor equipment. Movement of both the sensor equipment and the indenter can also be applied. The force is measured during the application of the indenter and forms the basis for the simulated force applied in the simulation. Measuring the force rather than applying a specifically defined force has proven appropriate, as defined applications are possible but more complex. The pressure value is typically the output signal of a barometric pressure sensor.

[0049] It should be noted that it is not necessary to perform force tests by simultaneously applying multiple indenters in order to prepare a feedforward neural network for correctly evaluating multiple forces. This can be done through simulations to train the transfer network, as mentioned above.

[0050] Simulations for training a transfer network can be performed using the same finite element model used for training a reconstructed network.

[0051] In simulations, the simulated forces and the structural and material properties of the finite element model typically form the basis for calculations performed by the finite element model. In particular, the simulated forces lead to calculated simulated force maps and calculated virtual sensor point values. Therefore, the finite element model is used to calculate virtual sensor point values ​​corresponding to the forces actually applied to the measurement surface.

[0052] A simulated force may, in particular, correspond to an actually measured force. For example, a simulated force may have at least one of the following: the same component, the same absolute value, and the same direction (as the measured force). In particular, a simulated force may have an integral over the contact area of ​​a simulated indenter to which the simulated force is applied to a simulated measurement surface, and may be equal to, or have a predefined relationship with, the integral of at least one of the measured force or the actual force over the actual contact area and the measured force. This may relate, for example, to at least one of the force's amplitude and direction. Alternatively, a predefined variation between the measured force and the simulated force can be used and can also be considered a corresponding force.

[0053] Position can be obtained, for example, by measurement, image recognition using a camera, or by calculation from mechanical variables when performing force tests. Simulated forces may be applied, in particular, to the same positions on the simulated measurement surface as the positions on the actual measurement surface where the actual force is applied. This allows for a good correspondence between experiment and simulation.

[0054] When training a transfer network, both experimental and simulation data can be used. Such simulation data includes, in particular, pressure values ​​from a barometric pressure sensor and values ​​from a corresponding virtual sensor point in the simulation.

[0055] In one implementation, force tests used to train the transfer network are performed using multiple indenters, each with its own indenter shape. The shape can relate particularly to the part of the indenter that contacts the measurement surface during the force test. Thus, the indenter is the object used in the force test to determine the force applied to the measurement surface. In particular, multiple force tests can be performed using one of the group of indenter shapes in each force test. Typically, only one indenter is used in each force test.

[0056] In one implementation, the indenter shape is selected from the group comprising at least a tip, a circle, a triangular cross-section, a square cross-section, a hemisphere, a cube, and a cylinder. Such indenter shapes have proven appropriate because they correspond to the typical shapes of objects that come into contact with the measurement surface in the application. Using such different indenter shapes greatly improves the training of the transfer network to reconstruct the corresponding or similar shapes applied to the measurement surface. Note that each of the mentioned shapes can be used, only one of the mentioned shapes can be used, or a selection can be used from the mentioned shapes. Alternatively, or in addition, other shapes can be used.

[0057] In one implementation, the simulation is performed with simulated forces based on simulated indenters, each having a simulated indenter shape corresponding to the actual indenter shape used in the corresponding force test. This ensures an optimal correspondence between the force test and the simulation, allowing the transfer network to be ideally trained.

[0058] In one implementation, the transfer network was trained using multiple different indenter shapes. This allowed the transfer network to be trained to distinguish between multiple forces generated by different indenter shapes. Typically, the different indenter shapes are distributed across multiple force tests, as only one indenter is applied to each force test.

[0059] In one implementation, the transfer network was trained using multiple indenters of different sizes. Furthermore, instead of using, for example, different shapes, this allows the transfer network to be trained to distinguish between indenters or other objects applying forces of different sizes. For example, different sized contact portions can be used on the measurement surface.

[0060] In one implementation, the transfer network was trained with indenters subjected to each shear force for at least a portion of the force tests used to train the transfer network. This allows the transfer network to be trained to distinguish between various shear forces applied to the measurement surface. In particular, multiple force tests can be performed using different shear forces or shear force components.

[0061] In one implementation, the measured force includes a normal force component, a first shear force component, and a second shear force component, respectively. Therefore, the measured force provides information about these components. It should be noted that in typical prior art implementations, the shear force could not be measured. However, it has been shown that using such a simulated force vector with the mentioned components to train a force test transfer network allows for the reconstruction of the shear force in addition to the normal force. This provides valuable additional information in several applications, such as robotic applications controlling the end effector of a robot.

[0062] In this implementation, the first shear force component of the measured force corresponds to the first shear force, and the second shear force component corresponds to the second shear force. In particular, the first shear force is perpendicular to the second shear force. This provides easily usable information based on the perpendicular direction of the shear force.

[0063] Please note that the measured force may, alternatively, contain around three components.

[0064] The measured force may be expressed in a global coordinate system. Alternatively, it may be expressed as multiple shear forces, where the normal component is locally perpendicular to the measurement surface at the contact point, and at least one of the normal components and the other being perpendicular to each other. This is considered equivalent because the components in another coordinate system can be calculated using coordinate transformations.

[0065] In one implementation, the transfer network was trained using multiple forces with different shear force components. This allows the transport network to be trained to distinguish between various shear forces applied to the measurement surface. The shear force can vary between different components of the force used in one simulation or in multiple different simulations.

[0066] In one implementation, the transition network was trained using multiple forces with different normal force components. This allows the transition network to be trained to distinguish between various normal forces applied to the measurement surface. Normal forces can differ particularly between various force tests and corresponding simulations.

[0067] In one implementation, the force applied by the indenter is measured using a force sensor placed inside or next to the indenter. Such a force sensor can measure the force applied to the measurement surface by the indenter. In particular, the three components of the force may be measured, for example, as discussed above. Placing the force sensor next to the indenter may include positioning it in at least one of the following: contact with the indenter and positioning it between the indenter and the object to which it is attached.

[0068] In one implementation, each simulated force vector comprises a normal force component, a first shear force component, and a second shear force component. These can specifically correspond to measured forces. Therefore, the simulated forces can be used in simulations corresponding to the forces actually applied in force tests.

[0069] In one implementation, the feedforward neural network directly maps the barometric pressure sensor to a force map. This can be seen as an alternative implementation to splitting the feedforward neural network into a transfer network and a reconstruction network. In particular, this implementation does not use mapping from pressure values ​​to virtual sensor point values. Instead, there is only one trained neural network that directly maps pressure values ​​to a force map.

[0070] For example, to properly train the transfer network, force tests and corresponding simulations can be performed between 20 and 100 cycles, or between 50 force tests and corresponding simulations.

[0071] As further examples, at least one of the following can be performed: at least 20 force tests, at least 50 force tests, at least 100 force tests, at least 500 force tests, at least 1,000 force tests, at least 2,000 force tests, or at least 10,000 force tests and at least one of up to 500 force tests, up to 1,000 force tests, up to 2,000 force tests, up to 10,000 force tests, or up to 50,000 force tests. However, other numbers may also be used.

[0072] Force tests can be performed in a way that measures the force, even if it is not predetermined in each case. Different apparatus variables can be used to obtain different forces.

[0073] In one implementation, the feedforward neural network may be trained by the following steps, which are performed before force inference: - A step of performing multiple force tests on a sensor device, wherein each force test is: The process involves applying a force to a position on the measurement surface of the sensor device using a single indenter, and simultaneously measuring the force applied by the indenter. A step including simultaneously measuring pressure values ​​using a barometric pressure sensor, - A step in which, for each force test, a corresponding simulation is performed using a finite element model of the sensor device, Each simulation involves applying simulated forces to a simulated measurement surface of a finite element model, thereby calculating a simulated force map on the simulated measurement surface. The simulated force map includes multiple simulated force vectors. The simulated force corresponds to the measured force and is applied to a position on the simulated measurement surface that corresponds to the position on the measurement surface; the steps to be performed are: - A step of training a feedforward neural network using measured pressure values ​​and corresponding calculated simulated force maps. That is the case.

[0074] Such training can be performed even when virtual sensor point values ​​are not used. This can be used, for example, in implementations where a feedforward neural network directly maps pressure values ​​to a force map, as discussed above. However, it can also be used in implementations where a feedforward neural network is split into a transfer network and a reconstruction network, as discussed above, in addition to training the transfer network and reconstruction network separately.

[0075] For details on force testing and simulation, refer to the above-mentioned descriptions regarding training the transfer network and training the reconstructed network.

[0076] In one implementation, force tests for training a feedforward neural network are performed using multiple indenters, each with its own distinct shape. The shape may relate particularly to the part of the indenter that contacts the measurement surface during the force test. Thus, an indenter is an object used in force tests to apply force to a measurement surface.

[0077] In one implementation, the indenter shape is selected from the group comprising at least a tip, a circle, a triangular cross-section, a square cross-section, a hemisphere, a cube, and a cylinder. Such indenter shapes have been proven appropriate because they correspond to the typical shapes of objects that come into contact with the measurement surface in the application. Using such different indenter shapes significantly improves the training of the feedforward neural network to reconstruct the corresponding or similar shapes applied to the measurement surface. Note that each of the mentioned shapes can be used, only one of the mentioned shapes can be used, or a selection of the mentioned shapes can be used. Alternatively, or in addition, other shapes can be used.

[0078] In one implementation, the simulation is performed with simulated forces based on simulated indenters, each having a simulated indenter shape corresponding to the actual indenter shape used in the corresponding force test. This ensures an optimal correspondence between the force test and the simulation, allowing for ideal training of a feedforward neural network.

[0079] In one implementation, the feedforward neural network was trained using multiple different indenter shapes. This allowed the feedforward neural network to be trained to distinguish between the forces generated by different indenter shapes.

[0080] In one implementation, a feedforward neural network was trained using multiple indenters of different sizes. Furthermore, by using different sizes instead of different shapes, for example, the feedforward neural network can be trained to distinguish between indenters or other objects applying force based on their size. For example, different sized contact points can be used on the measurement surface.

[0081] In one implementation, a feedforward neural network was trained by indenters to which each shear force was applied, for at least a portion of the force tests used to train the feedforward neural network. This allows the feedforward neural network to be trained to distinguish between different shear forces applied to the measurement surface. In particular, multiple force tests can be performed using different shear forces or shear force components.

[0082] In one implementation, each measured force has a normal force component, a first shear force component, and a second shear force component. Therefore, the measured force provides information about these components. Note that in typical prior art implementations, shear force could not be measured. However, it has been shown that when such a simulated force vector with the above components is used to train a feedforward neural network in a force test, shear force can be reconstructed in addition to normal force. This provides valuable additional information in several applications, such as robotic applications controlling the end effector of a robot.

[0083] In this implementation, the first shear force component of the measured force corresponds to the first shear force, and the second shear force component corresponds to the second shear force. In particular, the first shear force is perpendicular to the second shear force. This provides easily usable information based on the perpendicular direction of the shear force.

[0084] Please note that the measured force may, alternatively, contain around three components.

[0085] In one implementation, a feedforward neural network was trained using multiple forces with different shear force components. This allows the feedforward neural network to be trained to distinguish between different shear forces applied to the measurement surface. The shear force can vary among the various components of the force used in one or more force tests.

[0086] In one implementation, a feedforward neural network was trained using multiple forces with different normal force components. This allows the feedforward neural network to train itself to differentiate different normal forces applied to the measurement surface. Normal forces can differ particularly between different force tests.

[0087] In one implementation, the force applied by the indenter is measured using a force sensor placed inside or next to the indenter. Such a force sensor may measure the force applied by the indenter to the measurement surface. In particular, three force components may be measured, for example, as discussed earlier. Placing the force sensor next to the indenter may include positioning by at least one of contacting the indenter and being positioned between the indenter and the object to which it is attached.

[0088] In one implementation, each simulated force vector includes a normal force component, a first shear force component, and a second shear force component. These can specifically correspond to measured forces. Therefore, the simulated forces can be used for simulations that correspond to the forces actually applied in force tests.

[0089] The following sections will describe aspects related primarily to the actual processing steps of force reasoning, rather than training.

[0090] In one implementation, the pressure values ​​on which the calculated force map is based are read simultaneously or at predetermined intervals. This ensures that all pressure values ​​are associated with the same application of force.

[0091] In a typical implementation, the force map is mm 2 (1×10 -6 m 2 ) A force vector of at least 0.25 per mm 2 A force vector of at least 0.5 per mm 2 A force vector of at least 0.75 per mm 2 A force vector of at least 1 per unit, mm 2 A force vector of at least 1.5 per unit, or mm 2 Each unit has at least two force vectors.

[0092] In a typical implementation, the force map is in mm 2 A maximum force vector of 0.25 per hit, in mm 2 A maximum force vector of 0.5 per hit, in mm 2 A maximum force vector of 0.75 per hit, in mm 2 A maximum force vector of 1 per hit, in mm 2 A maximum force vector of 1.5 per hit, or in mm 2 Includes a maximum force vector of 2 per hit.

[0093] Force vectors of such density provide sufficient resolution and can be obtained with widely available computing power, and have proven to be suitable for typical applications. Each low value can be combined with each high value to form appropriate intervals. Also, force vectors of other densities can be used.

[0094] In a typical implementation, the force map includes at least 500, at least 1000, or at least 2000 force vectors. In a typical implementation, the force map includes a maximum of 1000, a maximum of 2000, a maximum of 3000, or a maximum of 4000 force vectors. Such implementation forms can be used, for example, when the sensor device is at the tip of a robot of approximately human size.

[0095] In a preferred implementation, each force vector includes a normal force component, a first shear force component, and a second shear force component. Thereby, a force map providing appropriate three-dimensional information is obtained.

[0096] In particular, the first shear force component may correspond to the first shear force, and the second shear force component may correspond to the second shear force. The first shear force may be perpendicular to the second shear force in particular. This enables appropriate shear force information of the applied force provided by the force map.

[0097] One implementation further comprises reading a temperature value from a barometric pressure sensor and providing temperature information or a temperature map of the sensor device based on the temperature value. This provides additional temperature information that can be used, for example, in robot control applications. For example, the temperature measurement function present in the barometric pressure sensor can be used for this purpose.

[0098] When this method includes both force maps and simulated force maps, it should be noted that typically the force map relates to a sensor device and the simulated force map relates to a finite element model. A given description for one of these force maps is typically applicable to both of them.

[0099] The following describes individual methods for training the network. These methods are not part of the force inference method, but are performed separately to train the network. For details on each method, refer to the above descriptions regarding network training and force inference methods to avoid repetition.

[0100] The present invention relates to a method for training a reconstructed network. -Here, the reconstruction network maps the virtual sensors of the finite element model of the sensor device to a force map. The sensor device comprises a plurality of barometric pressure sensors and a tracking layer that covers these barometric pressure sensors and provides a measurement surface. A force map contains multiple force vectors. -Here, each virtual sensor has one or more virtual sensor points, and each virtual sensor point has a value. -Here, the reconstructed network is trained in the following steps. - A step of performing multiple simulations in a finite element model, Each simulation involves simultaneously applying one or more simulated forces to the simulated measurement surface of a finite element model, thereby calculating a simulated force map on the simulated measurement surface. The simulated force map includes multiple simulated force vectors, and steps, The steps include: calculating the corresponding virtual sensor point values ​​using a finite element model; - A step of training the reconstructed network using the calculated simulated force map and the corresponding calculated virtual sensor point values. That is the case.

[0101] In one implementation, the simulated force applied to the simulated measurement surface is generated based on each simulated indenter, which has a simulated indenter shape.

[0102] In one implementation, the simulated indenter shape is selected from the group comprising at least a tip, a circle, a triangular cross-section, a square cross-section, a hemisphere, a cube, and a cylinder.

[0103] In one implementation, the reconstruction network is trained using multiple different simulated indenter shapes.

[0104] In one implementation, the reconstruction network is trained using simulated indenters of multiple sizes.

[0105] In one implementation, the reconstruction network is trained by at least part of a simulation that involves the simultaneous application of simulated forces generated based on two or more simulated indenters.

[0106] In one implementation, the reconstruction network is trained by at least part of a simulation that involves the application of simulated forces generated based on only one simulated indenter.

[0107] In one implementation, each vector of the simulated force includes a normal force component, a first shear force component, and a second shear force component.

[0108] In one implementation, the first shear force component of the simulated force vector corresponds to the first shear force, the second shear force component corresponds to the second shear force, and the first shear force is perpendicular to the second shear force.

[0109] In one implementation, the reconstructed network is trained using multiple simulated forces with different shear force components.

[0110] In one implementation, the reconstruction network is trained using multiple simulated forces with different normal force components.

[0111] In one implementation, the reconstructed network is used in the manner described above with respect to the use of the transfer network and the reconstructed network.

[0112] In each implementation, - The force map is mm 2 (1×10 -6 m 2 ) A force vector of at least 0.25 per mm 2 A force vector of at least 0.5 per mm 2 A force vector of at least 0.75 per mm 2 A force vector of at least 1 per unit, mm 2 A force vector of at least 1.5 per unit, or mm 2 It has at least 2 force vectors per unit, - The force map is mm 2 Maximum force vector of 0.25 per unit, mm 2 Maximum force vector of 0.5 per unit, mm 2 Maximum force vector of 0.75 per unit, mm 2 Maximum force vector of 1 per unit, mm 2 A maximum force vector of 1.5 per unit, or mm 2 It has a maximum force vector of 2 per unit and It is at least one of the following.

[0113] In one implementation, each force vector includes a normal force component, a first shear force component, and a second shear force component.

[0114] 1. In implementation, - The first shear force component corresponds to the first shear force, and the second shear force component corresponds to the second shear force. - The first shear force is perpendicular to the second shear force.

[0115] The same may be true for the simulated force map and its simulated force vectors.

[0116] This invention relates to a method for training a transfer network. -Here, the transfer network maps the barometric pressure sensor of the sensor device to multiple virtual sensors in the finite element model of the sensor device, The sensor device, Multiple barometric pressure sensors, A follow-up layer that covers the pressure sensor and provides a measurement surface Equipped with, -Here, each virtual sensor has one or more virtual sensor points, and each virtual sensor point has a value. -Here, the transfer network is trained in the following steps. -A step of performing multiple force tests on a sensor device, Each force test involves applying a force using a single indenter to a position on the measurement surface of the sensor device, simultaneously measuring the force applied by the indenter, and simultaneously measuring the pressure value using a barometric pressure sensor. - A step in which a corresponding simulation is performed using a finite element model for each force test, Each simulation includes applying simulated forces to a simulated measurement surface of a finite element model, thereby calculating a simulated force map on the simulated measurement surface. The simulated force map contains multiple simulated force vectors. The simulated force is applied to a position on the simulated measurement surface that corresponds to the measured force and the position on the measurement surface, in the following steps: The steps include: calculating the corresponding virtual sensor point values ​​using a finite element model; - A step of training a transfer network using the measured pressure value and the corresponding calculated virtual sensor point value. That is the case.

[0117] In one implementation, the force test used to train the transfer network is performed using multiple indenters, each with a different indenter shape.

[0118] In one implementation, the indenter shape is selected from the group comprising at least a tip, a circular shape, a triangular cross-section, a square cross-section, a hemisphere, a cube, and a cylinder.

[0119] In one implementation, the simulation is performed with multiple simulated forces based on multiple simulated indenters, each having a simulated indenter shape that corresponds to the actual indenter shape used in the corresponding force test.

[0120] In one implementation, the transfer network is trained using multiple different indenter shapes.

[0121] In one implementation, the transfer network is trained using multiple indenters of different sizes.

[0122] In one implementation, the transfer network is trained with multiple indenters, each with its own shear force, for at least a portion of the force tests used to train the transfer network.

[0123] In one implementation, each measured force has a normal force component, a first shear force component, and a second shear force component.

[0124] In one implementation, the first shear force component of the measured force corresponds to the first shear force, and the second shear force component corresponds to the second shear force. Furthermore, the first shear force is perpendicular to the second shear force.

[0125] In one implementation, the transfer network is trained using multiple forces with different shear force components.

[0126] In one implementation, the transfer network is trained using multiple forces with different normal force components.

[0127] In one implementation, the force applied by the indenter is measured using a force sensor located inside or next to the indenter.

[0128] In one implementation, each vector of the simulated force comprises a normal force component, a first shear force component, and a second shear force component.

[0129] In one implementation, the transfer network is used in the manner described above with respect to the use of the transfer network and the reconstructed network.

[0130] This invention relates to a method for training a feedforward neural network. -Here, the feedforward neural network calculates a force map on the measurement surface of the sensor device based on the pressure value of the barometric pressure sensor. The sensor device is Multiple barometric pressure sensors, The system comprises a conforming layer that covers the pressure sensor and provides a measurement surface, A force map contains multiple force vectors. - Here, the feedforward neural network is trained in the following steps. - A step of performing multiple force tests on the sensor device, Each force test comprises applying a force with a single indenter at a position on the measurement surface of the sensor device, simultaneously measuring the force applied by the indenter, and simultaneously measuring the pressure value with a barometric pressure sensor. - A step in which, for each force test, a corresponding simulation is performed using a finite element model of the sensor device, Each simulation includes applying simulated forces to a simulated measurement surface of a finite element model, thereby calculating a simulated force map on the simulated measurement surface. The simulated force map includes multiple simulated force vectors, The simulated force is applied to a position on the simulated measuring surface that corresponds to the measured force and the position on the measuring surface, and the steps are as follows: - A step of training a feedforward neural network using measured pressure values ​​and corresponding calculated simulated force maps. That is the case.

[0131] In one implementation, the force test used to train the feedforward neural network is performed using multiple indenters, each with its own indenter shape.

[0132] In one implementation, the indenter shape is selected from the group comprising at least a tip, a circular shape, a triangular cross-section, a square cross-section, a hemisphere, a cube, and a cylinder.

[0133] In one implementation, the simulation is performed with multiple simulated forces based on multiple simulated indenters, each having a simulated indenter shape that corresponds to the actual indenter shape used in the corresponding force test.

[0134] In one implementation, a feedforward neural network is trained using multiple different indenter shapes.

[0135] In one implementation, a feedforward neural network is trained using multiple indenters of different sizes.

[0136] In one implementation, a feedforward neural network is trained with multiple indenters, each with a different shear force, for at least a portion of the force tests used to train the feedforward neural network.

[0137] In one implementation, each measured force comprises a normal force component, a first shear force component, and a second shear force component, respectively.

[0138] In one implementation, of the multiple forces measured, the first shear force component corresponds to the first shear force, the second shear force component corresponds to the second shear force, and the first shear force is perpendicular to the second shear force.

[0139] In one implementation, the feedforward neural network is trained using multiple forces with different shear force components.

[0140] In one implementation, a feedforward neural network is trained using multiple forces with different normal force components.

[0141] In one implementation, the force is measured using a force sensor located inside or next to the indenter.

[0142] In one implementation, each vector of the simulated force has a normal force component, a first shear force component, and a second shear force component.

[0143] In one implementation, among the multiple vectors of the simulated force, the first shear force component corresponds to the first shear force, the second shear force component corresponds to the second shear force, and the first shear force is perpendicular to the second shear force.

[0144] In one implementation, a feedforward neural network is used in the force inference method described above.

[0145] In each implementation, - The force map is mm 2 (1×10 -6 m 2 ) A force vector of at least 0.25 per mm 2 A force vector of at least 0.5 per mm 2 A force vector of at least 0.75 per mm 2 At least one force vector per mm 2 A force vector of at least 1.5 per unit, or mm 2 Each unit must have at least two force vectors. - The force map is mm 2 Maximum force vector of 0.25 per unit, mm 2 Maximum force vector of 0.5 per unit, mm 2 Maximum force vector of 0.75 per unit, mm 2 Maximum force vector of 1 per unit, mm 2 A maximum force vector of 1.5 per unit, or mm 2 Maximum win 2 It has a force vector of 2. It is at least one of the above.

[0146] In one implementation, each force vector includes a normal force component, a first shear force component, and a second shear force component.

[0147] In one implementation, the first shear force component corresponds to the first shear force, and the second shear force component corresponds to the second shear force. The first shear force is perpendicular to the second shear force.

[0148] Details of sensor devices to which the methods disclosed herein can be applied are described below. Details or descriptions of such sensor devices given herein may be further referenced and applied accordingly.

[0149] In particular, in the method disclosed herein, the sensor device may be a force-sensing sensor device, and the sensor device is - Flexible circuit board and, - Multiple pressure sensors mounted on a flexible circuit board, -A rigid core that surrounds a flexible circuit board and to which the flexible circuit board is attached, wherein the rigid core, together with the pressure sensor protruding away from the rigid core, is at least partially covered by the flexible circuit board. - A follow-up layer that covers the pressure sensor and provides a measurement surface. It is equipped with.

[0150] However, it should be noted that the force reasoning and training concepts disclosed herein can also be applied to other sensor devices. This is particularly relevant to the use of at least one of different indenter sizes, different indenter shapes, different shear forces, and different shear force components. Such concepts can be generalized.

[0151] In one implementation, the rigid core is dome-shaped.

[0152] In one implementation, the rigid core has multiple planar sections, where each pressure sensor is located on one of these planar sections.

[0153] In one implementation, the follow-up layer contains or consists of plastic material or rubber.

[0154] According to the practice, the plastic material is a thermoplastic, elastomer, thermoplastic elastomer, or thermosetting resin.

[0155] In one implementation, the tracking layer relays the force applied to the measurement surface to at least a portion of the barometric pressure sensor.

[0156] In one implementation, the pressure sensor is connected by a conductive path on a flexible circuit board.

[0157] In one implementation, the flexible circuit board is shaped like an asterisk.

[0158] In one implementation, the flexible circuit board has multiple arms connected at its center.

[0159] In each implementation, the barometric pressure sensor is positioned at least one of the following distances: at least 1 mm, at least 2 mm, at least 3 mm, at least 4 mm, or at least 5 mm, and at a maximum distance of 1 mm, at a maximum of 2 mm, at a maximum of 3 mm, at a maximum of 4 mm, or at a maximum of 5 mm.

[0160] Depending on the implementation, the sensor device is located at least one of the robot's end-effector and the robot's operating element.

[0161] In one implementation, the rigid core is a three-dimensionally fabricated component.

[0162] The present invention further relates to a force inference module for force inference of a force-sensing sensor device. This force inference module is configured to perform a method as disclosed herein. With respect to the method, all embodiments and variations are applicable.

[0163] The present invention relates to a sensor device for detecting force, and the sensor device comprises one, some, or all of the following: - Flexible circuit board and, - Multiple pressure sensors mounted on a flexible circuit board, -A rigid core that surrounds a flexible circuit board and to which the flexible circuit board is attached, wherein the rigid core, together with the pressure sensor protruding away from the rigid core, is at least partially covered by the flexible circuit board. - A follow-up layer that covers the pressure sensor and provides a measurement surface, - Force inference module according to the present invention That is the case.

[0164] With respect to sensor devices equipped with a force inference module, all embodiments and variations of the force inference module, sensor device, and its components may be applied, particularly as described herein.

[0165] Further inventive aspects are described below. These aspects, either alone or in combination with other features disclosed herein, may also be considered as separate inventive aspects and may be the subject of claims.

[0166] The present invention relates to a sensor device for detecting force. The sensor device comprises a flexible circuit board. The sensor device comprises a plurality of barometric pressure sensors mounted on the flexible circuit board. The sensor device comprises a rigid core to which the flexible circuit board is wrapped and mounted, so that the flexible circuit board at least partially covers the rigid core with barometric pressure sensors protruding from the rigid core. The sensor device further comprises a conforming layer that covers the barometric pressure sensors and provides a measurement surface.

[0167] Such sensor devices can be manufactured at low cost and provide high resolution.

[0168] It should be understood that flexible circuit boards are pliable when handled independently, especially before being mounted on a rigid core. Such flexible circuit boards are easy to manufacture and handle, reducing labor and costs. Barometric pressure sensors can be standard types used in many industrial or scientific applications. Therefore, standard barometric pressure sensors are very inexpensive. A barometric pressure sensor typically provides an output signal, i.e., a pressure value, that is proportional to the force applied to the sensor. This can be considered the definition of a barometric pressure sensor. In general, sensors of any pressure can be used.

[0169] A rigid core is typically made of a hard material such as plastic or metal. This provides stability to the sensor device, especially when force is applied. Thus, while the conforming layer may deform in response to the applied force, the rigid core absorbs the force and provides an immovable reference point.

[0170] The characteristic that the flexible circuit board at least partially covers the rigid core typically means that at least a portion of the rigid core is covered by the flexible circuit board. Typically, the rigid core may have a surface intended to be covered by the flexible circuit board, and the flexible circuit board may partially or completely cover this surface. In this way, the flexible circuit board may leave a portion of the rigid core's surface uncovered.

[0171] The flexibility of a flexible circuit board typically means that it can be easily bent, especially when it is separated and not yet mounted on a rigid core. For example, when separated, a flexible circuit board can behave like a piece of cloth or rubber.

[0172] Flexible circuit boards are typically attached to rigid cores using adhesive or screws. However, other means of attaching flexible circuit boards to rigid cores may also be used.

[0173] Typically, the pressure sensor is already mounted on the flexible circuit board before the flexible circuit board is attached to the rigid core.

[0174] A flexible circuit board may have multiple conductive paths, such as wires connecting a barometric pressure sensor, where the conductive paths are capable of supplying power and / or reading power. The use of a flexible circuit board is a highly efficient method for providing at least one of the power supply and / or reading functions of such a barometric pressure sensor mounted on a rigid core of individual shape, because the wires on the flexible circuit board automatically adapt to the required shape.

[0175] The tracking layer is, in particular, a layer that can deform in response to a force applied to the measurement surface. Such deformation is characteristic of the external force or of other variables such as the shape of the indenter or shear force. The tracking layer may have at least one of flexibility and elasticity so that it automatically recovers to a predetermined shape after the force is stopped being applied. The force applied to the measurement surface is typically relayed by the tracking layer to the barometric pressure sensors. In particular, for a typical force applied to the measurement surface, the tracking layer relays the force to the barometric pressure sensors in such a way that multiple barometric pressure sensors are affected by that force. Thus, a very high resolution can be obtained for detecting the force even if the barometric pressure sensors are spaced much wider apart than high-resolution sensors known in the prior art. This is due to the fact that more sophisticated force inference techniques can be used, for example, based on at least one of machine learning and artificial neural networks, as described in this application. In particular, one tracking layer may cover all barometric pressure sensors.

[0176] In one embodiment, the rigid core is dome-shaped. This is particularly suitable when the sensor device is at the end of a robot or another operating element. However, other shapes can also be used. For example, the sensor device can be adapted to design sensors for the legs, shins, thighs, or chest of a robot. The shape of the rigid core can be adjusted accordingly. For example, it may be planar, cylindrical, or random in shape. Typically, the shape is designed so that a contact force activates multiple barometric pressure sensors simultaneously, thereby localizing the force.

[0177] A rigid core may have multiple facets. In one implementation, each barometric pressure sensor, or at least a portion of a barometric pressure sensor, is positioned on at least one of the multiple facets. Therefore, the orientation of each barometric pressure sensor may be determined by the orientation of the respective facet on which it is positioned. This does not preclude the possibility of multiple barometric pressure sensors being positioned on their respective facets. Furthermore, there may be facets or other parts of the rigid core's surface on which no barometric pressure sensors are attached.

[0178] It should be noted that while the pressure sensor is placed on a flexible circuit board, the flexible circuit board typically conforms to the shape of the planar portion of the rigid core. Thus, the flexible circuit board forms its own planar portion.

[0179] In particular, multiple planar sections can be made to have different orientations, thereby enabling the measurement of forces in different directions.

[0180] The follow-through layer may contain or consist of plastic material or rubber. The plastic material may be, for example, thermoplastic, elastomer, thermoplastic elastomer, thermosetting, or similar material. Such materials have proven suitable for typical applications. However, other materials may also be used, particularly in the process of fabricating the follow-through layer described above.

[0181] The tracking layer may, in particular, relay the force applied to the measurement surface to at least some of the barometric pressure sensors. In particular, it may be configured to relay the force to multiple barometric pressure sensors for at least some or most of the measurement surface. This can improve the resolution when measuring force using electronic force inference techniques.

[0182] Barometric pressure sensors can be connected, particularly by conductive paths on flexible circuit boards. These conductive paths can be especially flexible, automatically adapting to the surface of a rigid core when the flexible circuit board is wrapped around it. Such conductive paths enable reliable and simple connections for barometric pressure sensors.

[0183] Barometric pressure sensors can be connected, particularly by conductive paths on flexible circuit boards. These conductive paths can be especially flexible, so they automatically adapt to the surface of a rigid core when the flexible circuit board is wrapped around it. Such conductive paths allow for reliable and simple connection of barometric pressure sensors.

[0184] The flexible circuit board may be asterisk-shaped in particular. Specifically, it may have multiple arms or spokes connected in the center. This allows it to enclose a flexible circuit board, particularly on a dome-shaped rigid core, as can be seen, for example, from the attached drawings.

[0185] Barometric pressure sensors can be placed at distances of at least 1 mm, at least 2 mm, at least 3 mm, at least 4 mm, or at least 5 mm. They may also be placed at distances of up to 1 mm, up to 2 mm, up to 3 mm, up to 4 mm, or up to 5 mm. The distance may be measured around the outside of the barometric pressure sensors. Two different values ​​may be combined to form an appropriate spacing.

[0186] In particular, the sensor device may be the end effector of the robot and at least one of the robot's operating elements. While this is a preferred application, it should be noted that, in principle, the sensor device can also be used for several other applications, particularly when it is necessary to measure force and when it is necessary to use a certain element for manipulation. Manipulation, in particular, means that the operating element (which may be identical to, for example, the sensor device) can grasp or capture an article to be manipulated and manipulate this article, for example, in terms of its position or orientation. While this is happening, the force can be measured using the sensor device. In general, sensor devices such as those disclosed herein may combine barometric pressure sensing technology with novel assembly methods. Furthermore, they may be combined with machine learning techniques to create high-resolution tactile sensors with a high level of robustness, such as those having a three-dimensional dome shape, as disclosed herein.

[0187] In one embodiment, the rigid core is a three-dimensionally fabricated part. This allows for variable and efficient manufacturing. However, other manufacturing methods (e.g., drilling, molding) can also be used.

[0188] For example, at least 5, at least 10, at least 15, at least 19, at least 20, at least 25, at least 30, at least 35, or at least 37 barometric pressure sensors can be used. The sensors may be arranged to cover the entire dome-shaped central core. Such an assembly may be placed in a mold and thereafter covered with a material (e.g., urethane) that provides a flexible outer surface to protect the sensors, while also enabling local pressure measurement. Typically, since the (multiple) barometric pressure sensors are separate elements, each barometric pressure sensor may be distinguishable from neighboring sensors at least visually and physically.

[0189] The present invention further relates to a method for fabricating a force-sensing sensor device. This method comprises the following steps: - The step of providing a flexible circuit board with multiple pressure sensors attached, - Steps include providing a rigid core, - A step of wrapping and attaching a flexible circuit board to the rigid core, wherein the flexible circuit board wraps and attaches the rigid core, such that the flexible circuit board at least partially covers the rigid core with the pressure sensor attached, which protrudes away from the rigid core. - A step of covering a pressure sensor with a tracking layer, thereby providing a measurement surface on the tracking layer and That is the case.

[0190] Such methods can be used, in particular, for the fabrication of the aforementioned sensor devices. It should be noted that all descriptions given regarding sensor devices can also be applied to methods for fabricating them. The same applies in the reverse direction (applying matters concerning fabrication methods to matters concerning configuration), as long as such descriptions are technically appropriate.

[0191] This method provides an inexpensive and efficient way to manufacture sensor devices, particularly sensor devices according to the present invention.

[0192] In one implementation, the pressure sensor is already mounted on the flexible circuit board when starting the method described. However, in alternative implementations, mounting the pressure sensor to the flexible circuit board may be part of the method described, as further explained below, for example.

[0193] A flexible circuit board can be attached to a rigid core by, for example, using adhesive, screws, clamps, or by any other means, securing the flexible circuit board so as to cover at least a portion of the surface of the rigid core that holds the flexible circuit board.

[0194] The barometric pressure sensor protrudes away from the rigid core, which improves how forces applied to the barometric pressure sensor due to forces applied to the measurement surface are actuated. Specifically, this means that the rigid core is in contact with one surface of the flexible circuit board, and the barometric pressure sensor is mounted (attached) to the other surface of the flexible circuit board.

[0195] When the conforming layer covers the pressure sensor, it typically covers multiple points on the flexible circuit board, particularly those outside the pressure sensor, and at least a portion of the rigid core. In particular, the conforming layer may directly contact the rigid core in surface areas covered by the conforming layer but not by the flexible circuit board. Covering by the conforming layer may be carried out in particular as further described below.

[0196] Preferably, covering the pressure sensor with a tracking layer comprises the following steps: - The step of placing a rigid core with a flexible circuit board attached into a mold, - The step of filling the mold with material at least partially so that the pressure sensor is covered with material, - A step to change the material into a follow-up layer and That is the case.

[0197] Such a method of coating with a conforming layer provides an easy and cost-effective manufacturing process. The mold may define the final shape of the measurement surface, in particular, such that the measurement surface of the conforming layer obtains a shape defined by the shape of the mold.

[0198] The mold may be partially filled with material or completely filled. This depends on which portion of the rigid core or the flexible circuit board attached to the rigid core needs to be covered by the conforming layer. In particular, the mold may be filled with material to such an extent that the flexible circuit board is completely covered by the material.

[0199] Changing the material to a conforming layer means that, for example, it can be a fluid that can be easily filled into a mold, thus allowing the use of an easy-to-handle material.

[0200] Changing (the materials) may involve, for example, the following steps: A step of degassing the material by placing a rigid core with a flexible circuit board covered in the material into a vacuum.

[0201] Therefore, for example, a material that is a fluid in a non-degassed state but forms a follow-up layer in a degassed state can be used.

[0202] Degassing can be performed, particularly at room temperature, for example, in the temperature range between 15°C and 25°C. During vacuum conditions, the temperature may rise compared to such values.

[0203] However, it should be noted that other techniques for forming follow-up layers may also be used.

[0204] Providing a flexible circuit board may involve one or both of the following steps. - A step of cutting out at least a portion of the flexible circuit board from a sheet, - A step of placing and mounting the pressure sensor on a flexible circuit board.

[0205] Therefore, the preparation of a flexible circuit board equipped with the pressure sensor may be part of the method. In an alternative embodiment, a flexible circuit board on which the pressure sensor is already mounted may be used.

[0206] In one embodiment, the rigid core is dome-shaped. This is particularly suitable when the sensor device is at the end of a robot or another operating element. However, other shapes can also be used. For example, the sensor device may be adapted to design a sensor for the leg, shin, thigh, or chest of a robot. The shape of the rigid core is tailored according to the sensor device. For example, it may be planar, cylindrical, or randomly shaped. Typically, the shape is designed so that a contact force activates multiple barometric pressure sensors simultaneously, thereby localizing the force.

[0207] The rigid core may have multiple planar sections. In one implementation, each barometric pressure sensor, or at least some of the multiple barometric pressure sensors, is positioned on at least one of the planar sections. Thus, the orientation of each barometric pressure sensor may be determined by the orientation of the respective planar section on which it is positioned. This does not preclude the possibility of more than one barometric pressure sensor being positioned on each planar section. There may also be planar sections or other parts of the rigid core surface on which no barometric pressure sensors are attached.

[0208] When a pressure sensor is placed on a flexible circuit board, it should be noted that the flexible circuit board typically conforms to the shape of the planar portion of the rigid core. Therefore, the flexible circuit board itself forms multiple planar portions.

[0209] In particular, multiple planar sections may be given different orientations, thereby enabling the measurement of forces in different directions.

[0210] The follow-through layer may contain or consist of plastic material or rubber. The plastic material may be thermoplastic, elastomer, thermoplastic elastomer, thermosetting, or similar material. Such materials have proven suitable for typical applications. However, other materials may also be used. In particular, they may be used in the process of fabricating the follow-through layer as described above.

[0211] The tracking layer may, in particular, relay the force applied to the measurement surface to at least some of the multiple barometric pressure sensors. Specifically, it may be configured to relay the force to more than one barometric pressure sensor for at least a portion or most of the measurement surface. This can improve the resolution when measuring the force using electronic force inference techniques.

[0212] The pressure sensor may be connected by conductive paths, particularly on a flexible circuit board. These conductive paths can be flexible, so that they naturally adapt to the surface of the rigid core when the flexible circuit board is wrapped around it. Such conductive paths allow for reliable and simple connection of the pressure sensor.

[0213] The flexible circuit board may have an asterisk shape in particular. Specifically, it may have multiple arms or spokes connected in the center. This allows the flexible circuit board to cover a rigid core, particularly a dome-shaped one, as can be seen, for example, in the attached drawings.

[0214] The barometric pressure sensors may be placed at a distance of at least 1 mm, at least 2 mm, at least 3 mm, at least 4 mm, or at least 5 mm. They may also be placed at a distance of up to 1 mm, up to 2 mm, up to 3 mm, up to 4 mm, or up to 5 mm. The distance may be measured around the outside of the barometric pressure sensors. Two different values ​​can be combined to form an appropriate spacing.

[0215] In particular, the sensor device may be at least one of the robot's end-effector and the robot's operating element. While this is a preferred application, it should be noted that, in principle, this sensor device can also be used for several other applications, particularly when it is necessary to measure force and when the element is used for manipulation. Manipulation means, in particular, that an operating element, which may be identical to the sensor device, can grasp or capture an article to be manipulated, or manipulate the article in terms of its position or orientation. While doing so, force can be measured using the sensor device (of this application). In general, the sensor devices disclosed herein may combine barometric pressure sensing technology with novel assembly methods. Furthermore, they may be combined with machine learning techniques to create high-resolution tactile sensors with a high level of robustness, such as those having a three-dimensional dome shape, as disclosed herein.

[0216] In particular, the rigid core may be manufactured in three dimensions (3D printing). This means that providing a rigid core may involve the step of manufacturing the rigid core in three dimensions. However, other manufacturing methods such as drilling or molding may also be used.

[0217] For example, at least 5, at least 10, at least 15, at least 19, at least 20, at least 25, at least 30, at least 35, or at least 37 barometric pressure sensors can be used. They can cover the entire dome-shaped central core. Such assemblies can be placed in a mold covered with a material (e.g., urethane) to provide a flexible outer surface to protect the sensors while also enabling localized pressure measurement.

[0218] Machine learning techniques can be used to leverage sensor data and provide super-resolution sensing of tactile interactions. Therefore, multiple barometric pressure sensors may behave as if there were actually more sensors. The machine learning algorithm can be integrated by first training an eigenmodel of the finger pad using a finite element method, and then correlating the actual physical barometer with the eigenmodel using transfer learning. A force distribution map (nodal forces with 3 degrees of freedom and local coordinate systems) can be predicted as a representation of touch impact and categorized into various operation scenarios such as holding, inversion detection, and twisting.

[0219] This approach enables high-resolution sensing around the entire circumference of the finger's outline, making it ideal for a variety of applications where the object's contact position cannot be predicted, such as those that change significantly. Furthermore, the hardware components used in the sensor device are significantly less expensive, especially compared to other sensors known in the prior art.

[0220] In one preferred embodiment, the sensor device according to the present invention, or a sensor device manufactured according to a method according to the present invention, further comprises an electronically controlled module configured to perform a method for inferring the force of the sensor device. This allows the force inference function to be integrated into the sensor device. The electronically controlled module may be located, for example, inside or on the rigid core, or separately from the rigid core.

[0221] In particular, the control module can be configured to perform a force inference method that provides a force map of the measurement surface and a force map comprising multiple force vectors. Such a force map can provide relevant information about the applied force that may have originated, for example, from an indenter or an object being manipulated that is pressing against the measurement surface.

[0222] The control module can be configured in particular to perform at least one of the methods of force reasoning and training, as will be further described below.

[0223] In a typical implementation, the force map is mm 2 A force vector of at least 0.25 per mm 2 A force vector of at least 0.5 per mm 2 A force vector of at least 0.75 per mm 2 A force vector of at least 1 per unit, mm 2 A force vector of at least 1.5 per unit, or mm 2 It may have at least two force vectors.

[0224] In a typical implementation, the force map is mm 2 Maximum force vector of 0.25 per unit, mm 2 Maximum force vector of 0.5 per unit, mm 2 Maximum force vector of 0.75 per unit, mm 2 Maximum force vector of 1 per unit, mm 2 A maximum force vector of 1.5 per unit, or mm 2 It may have a force vector of up to 2 per unit.

[0225] These values ​​have proven suitable for typical use cases. However, other values ​​can also be used.

[0226] In a typical implementation, a force map may have at least 500, at least 1000, or at least 2000 force vectors. In a typical implementation, a force map may have up to 1000, up to 2000, up to 3000, or up to 4000 force vectors. Such values ​​have proven particularly suitable for use cases involving sensor devices at the end of a robot. However, other values ​​can also be used.

[0227] Preferably, each force vector comprises a normal force component, a first shear force component, and a second shear force component. This provides information not only about the normal force but also about the shear force, which allows for better tuning of, for example, the robot's end-effector.

[0228] In particular, the first shear force component may correspond to the first shear force, and the second shear force component may correspond to the second shear force. The first shear force may be perpendicular to the second shear force. In particular, the shear force components may be perpendicular to each other.

[0229] In one implementation, the control module may be configured to read temperature values ​​from the barometric pressure sensor and provide temperature information or a temperature map of the sensor device based on the temperature values. This may provide additional information about the temperature distribution that can be used for control or monitoring purposes. In particular, multiple barometric pressure sensors may each have an integrated temperature measurement function that can be used for this purpose.

[0230] Further inventive aspects are described below. These aspects, either alone or in combination with other features disclosed herein, may also be considered as separate inventive aspects and may be the subject of claims.

[0231] Further aspects and advantages will be apparent to those skilled in the art from the following description of the enclosed drawings. These illustrate the following:

Brief Description of the Drawings

[0232] [Figure 1] Figure 1 shows the arrangement of the sensors. [Figure 2] Figure 2 shows the rigid core. [Figure 3] Figure 3 shows the flexible circuit board. [Figure 4] Figure 4 shows the rigid core with the flexible circuit board attached. [Figure 5] Figure 5 shows an exploded view of the mold. [Figure 6] Figure 6 shows the assembled state of the mold. [Figure 7] Figure 7 shows an exploded view of the rigid core covered with the flexible circuit board and the mold. [Figure 8] Figure 8 shows the state of covering the rigid core with the flexible circuit board and the pressure sensor of the mold. [Figure 9] Figure 9 shows a schematic diagram of the force inference. [Figure 10] Figure 10 shows the finite element model. [Figure 11] Figure 11 shows the arrangement of a plurality of different force elements. [Figure 12] Figure 12 shows the arrangement for the force test. [Figure 13] Figure 13 shows a flowchart for training the transfer network. [Figure 14] Figure 14 shows a flowchart for training the reconstruction network. [Figure 15] Figure 15 shows a flowchart for training the forward propagation neural network. [Figure 16] Figure 16 shows the force map.

Modes for Carrying Out the Invention

[0233] Figure 1 shows a sensor device 10 according to an embodiment of the present invention.

[0234] The sensor device 10 comprises a dome-shaped rigid core 100. The rigid core 100 is partially covered by a flexible circuit board 300 fixedly mounted on the rigid core 100. The flexible circuit board 300 is covered by a conforming layer 200.

[0235] Multiple pressure sensors 400 are applied to the flexible circuit board 300. They protrude away from the rigid core 100. A conforming layer 200 provides a measuring surface 210 to which force can be applied. Because the conforming layer 200 is flexible and elastic, the force applied to the measuring surface 210 causes local deformation of the measuring surface 210, and the conforming layer 200 relays these forces to at least some of the pressure sensors 400. Thus, the pressure sensors 400 can be used to evaluate force or applied force.

[0236] The flexible circuit board 300 comprises a plurality of planar sections. These planar sections correspond to planar sections structured on the rigid core 100, as shown in detail in Figure 2.

[0237] The flexible circuit board 300 has a central portion 305 from which, in this embodiment, six arms extend. This central portion 305 can be considered a planar portion. All the arms are shown in Figure 3. In Figure 1, only three of these arms are visible, namely the first arm 310, the second arm 320, and the third arm 330, and are indicated by reference numerals.

[0238] Each arm is divided into three planar sections; for example, the first arm 310 is divided into a first planar section 311, a second planar section 312, and a third planar section 313. The other arms are divided accordingly, where the planar sections 321, 322, 323, 331, 332, and 333 of the flexible circuit board 300 are shown in Figure 1.

[0239] In the current embodiment, each planar section holds one pressure sensor 400. The central section 305 also holds one pressure sensor 400. Note that other configurations are possible; for example, the planar sections may have more than one pressure sensor 400, or they may not have any at all.

[0240] It should be noted that the pressure sensors 400 are spaced apart from each other on the flexible circuit board 300. However, much finer resolution with respect to applied force can be achieved using the method described below.

[0241] Figure 2 shows the rigid core 100 in sections. The rigid core 100 consists of a total of six surface areas, of which the first surface area 110, the second surface area 120, and the third surface area 130 are visible and shown in Figure 3. Each surface area 110, 120, and 130 is divided into three planar sections. For example, the first surface area 110 is divided into the first planar section 111, the second planar section 112, and the third planar section 113. The other surface areas are divided accordingly, and planar sections 121, 122, 123, 131, 132, and 133 are visible in Figure 1. At the top of the rigid core 100, the central section 105 connects the multiple surface areas.

[0242] The planar portion of the rigid core 100 defines the planar portion of the flexible circuit board 300. In detail, the planar portions have different orientations, and the flexible circuit board 300 adapts to each orientation of the planar portion.

[0243] Figure 2 clearly shows that the rigid core 100 is dome-shaped, and the rigid core can be used, for example, in the fingertips of a robot.

[0244] Figure 3 separately shows a flexible circuit board 300 equipped with a pressure sensor S400. As already mentioned, the flexible circuit board 300 has six arms 310, 320, 330, 340, 350, and 360 that are connected to each other at the central part 305. In this embodiment, a total of 19 pressure sensors 400 are attached to the flexible circuit board 300. More or fewer pressure sensors can be used in other embodiments.

[0245] Please note that there are no flat surfaces in Figure 3. This is because these flat surfaces in Figure 3 are not essential features of the flexible circuit board 300. The flat surfaces of the flexible circuit board 300 shown in Figure 1 are rather a result of the flexible circuit board 300 being mounted on the rigid core 100 shown in Figure 2.

[0246] Each arm 310, 320, 330, 340, 350, and 360 is provided with holes 315, 325, 335, 345, 355, and 365, respectively, which can be used, for example, to fasten the flexible circuit board 300 to the rigid core 100 during manufacturing.

[0247] Figure 4 shows the flexible circuit board 300 of Figure 3 attached to the rigid core 100 of Figure 2. In this way, the planar portion of the flexible circuit board 300 is already formed by the flexible circuit board 300 acquiring the structure of the rigid core 100. The configuration shown in Figure 4 does not yet have the conforming layer 200 shown in Figure 1. How the conforming layer 200 and its measurement surface 210 are formed will be shown with reference to the following figures.

[0248] Figure 5 shows an exploded view of the mold 500. The mold 500 comprises a first part 510 and a second part 520. As shown in Figure 5, hollow interiors 530 are formed inside parts 510 and 520 such that when parts 510 and 520 are assembled, the hollow interiors 530 open only to the top of the mold 500. In addition, as shown in Figure 4, the mold 500 includes a top 540 to secure the placement of a rigid core to which a flexible circuit board is attached.

[0249] FIG. 6 shows the mold 500 in an assembled state. Thus, the hollow interior 530 opens only at the upper part of the mold 500, and the top 540 spans over the hollow interior 530.

[0250] FIG. 7 shows the mold 500, already described, in an arrangement of the flexible circuit board 300 and the rigid core 100 on which the pressure sensor 400 is mounted. FIG. 7 shows an exploded view, and FIG. 8 shows the same in an assembled state. In the state shown in FIG. 8, the rigid core 100 is attached to the top 540 of the mold, and the rigid core 100 projects from the top 540 into the hollow interior 530.

[0251] In the state shown in FIG. 8, a material, for example a plastic material, can be filled into the hollow interior 530 in a fluid form. This is easy to handle due to its fluid properties. The material can be filled into the hollow interior 530 such that the flexible circuit board 300 and the rigid core 100 are covered by the material up to a level corresponding to the position where the follower layer 200 is to cover the flexible circuit board 300 and the rigid core 100. The surface of the hollow interior 530 defines the measurement surface 210 in the final state.

[0252] After filling the material, the mold 500, the rigid core 100, the flexible circuit board 300 attached thereto, and the already filled material are placed in a vacuum chamber. The vacuum chamber is evacuated to degas the material. By degassing, the material is transformed into the follower layer 200, and the sensor device 10 shown in FIG. 1 is in a manufactured state.

[0253] The processing steps shown with respect to these figures are manufacturing steps of the sensor device 10 that require only a few specific components and are easy to execute. Thus, the cost can be significantly reduced as compared to much more expensive embodiments known in the prior art.

[0254] Figure 9 shows a schematic diagram of the sensor device 10 as described above, or a method for inferring the force of the sensor device 10. As already mentioned, the sensor device 10 comprises a plurality of barometric pressure sensors 400. Each of these barometric pressure sensors 400 generates its own pressure values ​​R1, R2, ... Rx as its output value, indicating the pressure detected by each barometric pressure sensor 400 at a position below the tracking layer 200.

[0255] Such pressure values ​​R form the input to a transition network TN, which is a neural network that maps the barometric pressure sensor 400 to multiple virtual sensors in the finite element model 10a of the sensor device 10. The virtual sensors will be further described below with reference to Figure 10. Each virtual sensor contains one or more virtual sensor points, each virtual sensor point having values ​​S1, S2, ..., Sx. This will also be further described below with reference to Figure 10.

[0256] The fact that the transition network TN maps pressure values ​​R to virtual sensor point values ​​S means that for each combination of pressure values ​​R it takes as input, the transition network TN provides a set of virtual sensor point values ​​S as output. This requires training the transition network TN, which can be done as specifically described herein.

[0257] The values ​​S1, S2, ..., Sx of the virtual sensor points form the input to the reconstruction network RN, which is a neural network that maps the virtual sensors of the finite element model 10a to a force map FM. The force map FM consists of several force vectors F1, F2, ..., Fx, where each force vector F in the force map FM contains three components: a normal force component and two normal shear force components. Thus, each force vector F gives the value and direction of the force applied at a particular point on the measurement surface 210. The force map FM is further described with reference to Figure 16.

[0258] The fact that the reconstruction network RN maps the values ​​S of virtual sensor points to the force map FM means that the reconstruction network RN provides a set of force vectors F as an output for each combination of virtual sensor point values ​​S it takes as input. This requires training of the reconstruction network RN, which can be done as specifically described herein.

[0259] The transfer network TN and the reconstruction network RN together form a feedforward neural network FFNN. This should be considered a neural network that maps the barometric pressure sensor 400 to the force map FM. And this is divided into two parts, as already explained.

[0260] Method T1 may be used to train a transfer network (TN). Method T2 may be used to train a reconstruction network (RN). Method T3 may be used to train an entire feedforward neural network (FFNN). Such methods are described further below.

[0261] The use of neural networks, or artificial intelligence as a generalization, allows for the extraction of far more information from barometric pressure sensors than direct force inference without AI could provide. In particular, the applied force can be assessed with a much higher resolution than the spacing of the barometric pressure sensors 400. Furthermore, additional information such as shear force, the number of indenters, and their positions can be extracted. Such information is contained in a force map FM calculated based on the pressure value R.

[0262] Figure 10 shows a finite element model 10a of the sensor device 10. This finite element model 10a is used in the force inference process described in Figure 9. Figure 10 shows the structural details of the sensor device 10, but since such finite element concepts rely on known techniques, specific details of the implementation of finite element calculations are not shown. In principle, since the finite element model 10a is an electronic representation of the actual sensor device 10, the behavior of the sensor device 10 can be simulated using the finite element model 10a.

[0263] All components of the sensor device 10 have corresponding components in the finite element model 10a, and the components in the finite element model 10a are denoted by the letter "a". The structural difference between the sensor device 10 and the finite element model 10a is that the barometric pressure sensor 400 of the sensor device 10 is replaced by a virtual sensor 400a of the finite element model 10a. Each virtual sensor 400a has one or more sensor points 410a. Here, an implementation is shown in which each virtual sensor 400a has 12 virtual sensor points 410a. Each virtual sensor point 410a has a value S, as already discussed with respect to Figure 9. However, any other number of virtual sensor points 410a can also be used for each virtual sensor 400a.

[0264] In this way, the simulated force 605a applied to the simulated measurement surface 210a of the finite element model 10a is relayed to the virtual sensor 400a and its virtual sensor point 410a by the finite element representation of the tracking layer 200, i.e., the simulated tracking layer 200a. Such relayed forces produce values ​​S at each virtual sensor point. This can be used to run a simulation that gives the value S at each virtual sensor point for each applied simulated force 605a or combination of simulated forces 605a.

[0265] Such simulated force 605a is applied by a simulated indenter 600a. Two of these simulated indenters 600a are shown as examples in Figure 10. Using these simulated indenters 600a, a simulated force can be applied to the simulated measurement surface 210a, and the value S of the virtual sensor point can be calculated using a standard finite element model method.

[0266] Data obtained from such simulations can be used to train the reconstruction network RN. Typically, multiple such simulations, for example 1,000 simulations or about 10,000 simulations, are used for training. These simulations are typically performed using different types of simulated indenters 600a, particularly those with at least one of different shapes and sizes, and with different numbers of simulated indenters 600a, for example, having at least one indenter 600a, two indenters 600a, and three indenters 600a. Such simulations can be performed by pure computer simulation and do not require complex experimental setups (configurations). This enables highly efficient and reliable training of the reconstruction network RN, providing far more capabilities for reconstructing force maps FM even when experimental capabilities are limited.

[0267] Figure 11 schematically shows four different shapes of indenters 600, which can be used as physical indenters 600 or simulated indenters 600a for use in experimental settings, as further described below with respect to Figure 12.

[0268] Figure 11a shows an indenter 600 with a flat shape at its contact portion with the measurement surface 210. Figure 11b shows an indenter 600 with a pointed contact portion. Figure 11c shows an indenter 600 with a hemispherical contact portion. Figure 11d shows an indenter 600 with the same type of contact portion as the indenter 600 shown in Figure 11c, but in a relatively smaller size. Using such different indenters 600 allows for optimization of neural network training with respect to such different shapes, meaning that the ability of a neural network trained with such different indenters 600 increases with respect to the reconstructed force applied by indenters 600 with different indenter shapes. To give another example, the force map FM reconstructed after applying a flat-shaped indenter 600 is different from the force map FM reconstructed after applying a hemispherical-shaped indenter 600.

[0269] Figure 12 shows an experimental setup 700 for conducting force tests. The experimental setup 700 has a base 710 on which a first mechanical arm 720 is mounted. A joint 730 is positioned on the first mechanical arm 720. A second mechanical arm 740 is fixed to the joint 730. The joint 730 can be used to actively move the second mechanical arm 740, where an electric drive (not shown) is used for such movement.

[0270] A sensor device 10, as described earlier, is positioned at the other end of the second mechanical arm 740. This is shown schematically here. The outer surface of the sensor device 10 is the measuring surface 210, as already described.

[0271] The experimental setup 700 further comprises a top 750 to which a force sensor 610 is attached. An indenter 600 is attached to the force sensor 610. A joint 730 can be used to press the sensor device 10 against the indenter 600, and during such a force test, a pressure value R is read from the barometric pressure sensor 400, and the force 605 applied to the measurement surface 210 by the indenter 600 is measured by the force sensor 610. Since the force sensor 610 measures three-dimensional forces, both normal force components and shear force components are measured. The three-dimensional force may be represented in a global coordinate system, or it may be represented by a normal component (normal force component) perpendicular to a point on the measurement surface 210, and typically two shear force components that are perpendicular to the normal component and typically perpendicular to each other. Using coordinate transformations, if multiple components are known in another coordinate system, those components can be calculated.

[0272] The position where the indenter 600 contacts the measurement surface 210 is observed by the camera 620. This makes it possible to calculate the coordinates of this position on the measurement surface 210 using image recognition. Alternatively, such a position can be calculated, for example, using machine variables.

[0273] The fact that the indenter 600 is stationary and the sensor device 10 is moved in the experimental setup 700 allows for the use of jointed setups known, for example, from 3D printers. However, it should be noted that force tests can be performed in alternative or different ways, for example, by moving the indenter 600 with the stationary sensor device 10, or by moving both the sensor device 10 and the indenter 600.

[0274] Data derived from such force tests can be used to train the neural network shown in Figure 9. Figure 9 is explained further below.

[0275] Figure 13 shows a schematic diagram of method T1 for training the transfer network TN.

[0276] In the first step T1_1, multiple force tests are performed, as described with respect to Figure 12. For such force tests, different indenters 600 are preferably used, having at least one of different shapes and sizes, where only one indenter 600 is used in each force test in the described implementation.

[0277] In step T1_2, multiple simulations are performed using the finite element model 10a, where one simulation is performed for each force test, and the force 605 measured by the force sensor 610 in the force test is used in the simulation corresponding to the application of the simulated force 605a. The position on the simulated measurement surface 210a is the same as the position on the measurement surface 210 in the force test, where such a position can be calculated, for example, from mechanical variables or derived from image recognition, as already described with reference to Figure 12. The shape of the simulated indenter 600a is the same as the shape of the actual indenter 600. The value S of the virtual sensor point for each force test is calculated by a standard finite element simulation based on the applied simulated force 605a.

[0278] In step T1_3, the transfer network TN is trained using data obtained from force tests and simulations, in particular, the pressure value R of the barometric pressure sensor 400 resulting from the force test and the calculated virtual sensor point value S resulting from the corresponding simulation are used for training.

[0279] Figure 14 shows the method T2 for training the reconstructed network RN.

[0280] In the first step T2_1, multiple simulations are performed using the finite element model 10a, where preferably multiple different numbers of indenters are used, and more preferably multiple different indenter shapes and sizes are used. In each simulation, a simulated force map FMa is calculated on the simulated measurement surface 210a, and the corresponding virtual sensor point value S is calculated.

[0281] Using this simulated force map FMa and virtual sensor point values ​​S, the transfer network TN is trained in step T2_2 and can reconstruct the force map from the virtual sensor point values ​​S.

[0282] Figure 15 shows method T3 for training the entire feedforward neural network FFNN.

[0283] In the first step T3_1, multiple force tests are performed as described with respect to Figure 12. These force tests provide the applied force 605 (measured by the force sensor 610), the corresponding position, and the measured pressure value R from the pressure sensor 400.

[0284] In the second step T3_2, multiple corresponding simulations are performed using the finite element model 10a of the sensor device 10. Each simulation applies a simulated force 605a to the simulated measurement surface 210a of the finite element model 10, which is located at the same position as the actual measurement surface 210, and performs multiple corresponding simulations using a simulated indenter 600a that has the same indenter shape as the actual indenter 600. As a result, a simulated force map FMa is calculated on the simulated measurement surface 210a.

[0285] In a further step T3_3, measured pressure values ​​R from the force test and corresponding simulated force maps FMa resulting from the simulation are used to train the entire feedforward neural network (FFNN), and in the implementation shown here, both the transfer network TN and the reconstruction network RN are trained.

[0286] It should be noted that the process described with respect to Figure 15 can also be used when only one neural network is used, i.e., when partitioning is not implemented in the transfer network TN and the reconstructed network RN. In the implementation shown in Figure 9, both the transfer network TN and the reconstructed network RN can be optimized by performing the method described with respect to Figure 15, in addition to the methods described with respect to Figures 13 and 14.

[0287] Figure 16 shows the sensor device 10 along with a schematic diagram of the force map FM. The force map FM includes multiple force vectors F positioned around the entire circumference of the measurement surface 210. While two force vectors F are shown in Figure 16, many more force vectors F can be used in a typical implementation. For example, mm 2 (=1×10 -6 m 2 One force vector F per unit can be used in an exemplary implementation.

[0288] Each force vector F has a normal force component F N And the first shear force component F S1 And the second shear force component F S2 This includes the normal force component F. N This gives the value of the normal force component of the applied force, i.e., the value of the component perpendicular to the local orientation of the measurement surface 210. Shear force component F S1 F S2This gives the value of the shear force applied to the measurement surface 210 at each point. The shear force is typically parallel to the local orientation of the measurement surface 210, typically perpendicular to each other and perpendicular to the normal force. This may particularly relate to the undeformed orientation of the measurement surface, which can define the direction of the force vector F, especially its normal component.

[0289] Therefore, each force vector F represents the strength and direction of the force applied to a specific point on the measurement surface 210. Such a force may be, for example, due to an indenter 600.

[0290] Other definitions of the force vector F can also be used. For example, it may be possible to evaluate only the normal force component, or note that (multiple) shear forces may have alternative definitions.

[0291] In the case of the simulated force map FMa, the simulated force vector Fa of the simulated force map FMa on the simulated measurement surface 210a has simulated components, for example, the normal force component F. N a and the first shear force component F S1 a and the second shear force component F S2 It is acceptable to have a. Such a simulated force map FMa is specifically calculated in simulations performed on a finite element model, as explained with respect to Figure 10.

[0292] The steps of the method of the present invention can be performed in a given order. However, they can be performed in a different order, as long as it is technically reasonable. In embodiments, the method of the present invention can be performed, for example, using a specific combination of steps, so that no further steps are performed. However, other steps, including steps not mentioned, can also be performed.

[0293] While the various features (of this invention) may be used or implemented independently of each other, it should be noted that, for clarity, for example, the features are described in combination within the claims and specification. Those skilled in the art will realize that such features can be combined with other features, or that there may be combinations of features that are independent of each other.

[0294] References within dependent claims may indicate preferred combinations of features, but do not preclude other combinations of features. [Explanation of symbols]

[0295] 10 Sensor device 100 rigid core 105 Central part 110 1st surface area 111 Plane section 112 Plane section 113 Plane section 120 2nd surface area 121 Plane section 122 Plane section 123 Plane section 130 Third surface area 131 Plane section 132 Plane section 133 Plane section 200 follower layers 210 Measurement surface 300 Flexible Circuit Boards 305 Central part 310 First Arm 311 Plane section 312 Plane section 313 Plane section 315 holes 320 Second Arm 321 Plane section 322 Plane section 323 Plane section 325 holes 330 Third Arm 331 Plane section 332 Plane section 333 Plane section 335 holes 340 Fourth Arm 345 holes 350 Fifth Arm 355 holes 360 Sixth Arm 365 holes 400 bar sensor 500 molds 510 Part 1 520 Part 2 530 Hollow interior 540 Top 600 indenter 605 Power 610 Force Sensor 620 Camera 700 Experimental settings 710 Bottom 720 First Mechanical Arm 730 joints 740 Second Mechanical Arm 750 Top 10a Finite Element Model 210a Simulation measurement surface 400a Virtual Sensor 410a Virtual sensor point 600a simulation indenter 605a The Power of Simulation Other reference symbols with the letter 'a' are components 10a of the finite element model. TN transfer network RN Reconstruction Network FFNN (Feedback Neural Network) Training method for T1 transfer networks Training method for T2 reconstruction networks Training method for T3 feedforward networks R pressure value S Virtual sensor point value FM Power Map F force vector FMa simulated training force map Fa: Simulated training force vector F N (the normal force component of the force vector) FS1 (The first shear force component of the force vector) F S2 (The second shear force component of the force vector) F N a (Normal force component of the simulated force vector) F S1 a (the first shear force component of the simulated force vector) F S2 a (the second shear force component of the simulated force vector)

Claims

1. A method for inferring the force of a force-detecting sensor device (10), The sensor device (10) comprises a plurality of pressure sensors (400) and a follow-up layer (200) that covers the pressure sensors (400) and provides a measurement surface (210), The method for inferring the aforementioned force is, The steps include reading the pressure value (R) from the atmospheric pressure sensor (400), A step of calculating a force map (FM) on the measurement surface (210) based on the pressure value (R) using a feedforward neural network (FFNN), wherein the force map (FM) includes a plurality of force vectors (F), A method for inferring force, which includes the following features.

2. The aforementioned feedforward neural network (FFNN) includes a transfer network (TN) and a reconstruction network (RN), The transfer network (TN) maps the pressure sensor (400) to a plurality of virtual sensors (400a) of the finite element model (10a) of the sensor device (10). The reconstruction network (RN) maps the virtual sensor (400a) of the finite element model (10a) to the force map (FM), Each of the virtual sensors (400a) comprises one or more virtual sensor points (410a) having a value (S). The method according to claim 1.

3. Before force inference, the reconstruction network (RN) performs the following steps: - A step (T2_1) in which a plurality of simulations are performed in the finite element model (10a), Each of the simulations comprises simultaneously applying one or more simulated forces (605a) to the simulated measurement surface (210a) of the finite element model (10a), thereby calculating a simulated force map (FMa) on the simulated measurement surface (210a). The simulated force map (FMa) includes a plurality of simulated force vectors (F), Step (T2_1) involves calculating the value (S) of the corresponding virtual sensor point using the finite element model (10a), - A step (T2_2) of training the reconstructed network (RN) using the calculated simulated force map (FMa) and the corresponding calculated virtual sensor point values ​​(S), The method according to claim 2, trained by

4. The method according to claim 3, wherein the simulated force (605a) applied to the simulated measurement surface (210a) is generated based on each simulated indenter (600a) having a simulated indenter shape.

5. The method according to claim 4, wherein the simulated indenter shape is selected from the group comprising at least a tip, a circle, a triangular cross-section, a square cross-section, a hemisphere, a cube, and a cylinder.

6. The method according to any one of claims 3 to 5, wherein the reconstruction network (RN) is trained using a plurality of different simulated indenter shapes.

7. The method according to any one of claims 3 to 6, wherein the reconstruction network (RN) is trained using a plurality of sized simulated indenters (600a).

8. The method according to any one of claims 3 to 7, wherein the reconstruction network (RN) is trained by at least a portion of a simulation that includes the simultaneous application of simulated forces (605a) generated based on two or more simulated indenters (600a).

9. The method according to any one of claims 3 to 8, wherein the reconstruction network (RN) is trained by at least a portion of a simulation which includes applying a simulated force (605a) generated based on only one simulated indenter (600a).

10. Each of the simulated force vectors (Fa) has a normal force component (F Ｎ a) and the first shear force component (F Ｓ１ a) and the second shear force component (F Ｓ２ a) The method according to any one of claims 3 to 9, comprising:

11. Among the simulated force vector (Fa), the first shear force component (F Ｓ１ a) corresponds to the first shear force, and the second shear force component (F Ｓ２ a) corresponds to the second shear force, The method according to claim 10, wherein the first shear force is perpendicular to the second shear force.

12. The method according to any one of claims 3 to 11, wherein the reconstruction network (RN) is trained using a plurality of simulated forces (605a) having different shear force components.

13. The method according to any one of claims 3 to 12, wherein the reconstruction network (RN) is trained using a plurality of simulated forces (605a) having different normal force components.

14. Before force inference, the transfer network (TN) performs the following steps: - A step (T1_1) in which a plurality of force tests are performed on the sensor device (10), Step (T1_1) comprises each force test comprising applying a force with one indenter (600) at a position on the measurement surface (210) of the sensor device (10), simultaneously measuring the force (605) applied by the indenter (600), and simultaneously measuring the pressure value (R) with the barometric pressure sensor (400), - Step (T1_2) for each of the aforementioned force tests, in which a corresponding simulation is performed using the finite element model (10a), Each simulation includes applying a simulated force (605a) to the simulated measurement surface (210a) of the finite element model (10a), thereby calculating a simulated force map (FMa) on the simulated measurement surface (210a). The simulated force map (FMa) includes a plurality of simulated force vectors (Fa), Step (T1_2) is performed, wherein the simulated force (605a) is applied to a position on the simulated measuring surface (210a) that corresponds to the measured force (605) and the position on the measuring surface (210), The steps include: calculating the value (S) of the corresponding virtual sensor point using the finite element model (10a); - A step (T1_3) of training the transition network (TN) using the measured pressure value (R) and the corresponding calculated virtual sensor point value (S), The method according to any one of claims 2 to 13, trained in [the specified method].

15. The method according to claim 14, wherein the force test for training the transfer network (TN) is performed using a plurality of indenters (600), each indenter having its own indenter shape.

16. The method according to claim 15, wherein the indenter shape is selected from the group comprising at least a tip, a circular shape, a triangular cross-section, a square cross-section, a hemisphere, a cube, and a cylinder.

17. The method according to claim 15 or 16, wherein the simulation is performed by a simulated force (605a) based on a simulated indenter (600a) having a simulated indenter shape corresponding to the actual indenter shape used in the corresponding force test.

18. The method according to any one of claims 14 to 17, wherein the transfer network (TN) is trained using a plurality of different indenter shapes.

19. The method according to any one of claims 14 to 18, wherein the transfer network (TN) is trained using a plurality of indenters (600) of different sizes.

20. The method according to any one of claims 14 to 19, wherein the transition network (TN) is trained by indenters (600) to which each shear force is applied for at least a portion of a force test to train the transition network (TN).

21. The method according to any one of claims 14 to 20, wherein each measured force (605) includes a normal force component, a first shear force component, and a second shear force component.

22. In the measured force (605), the first shear force component corresponds to the first shear force, and the second shear force component corresponds to the second shear force. The method according to claim 21, wherein the first shear force is perpendicular to the second shear force.

23. The method according to claim 21 or 22, wherein the transition network (TN) is trained using a plurality of forces (605) having different shear force components.

24. The method according to any one of claims 21 to 23, wherein the transfer network (TN) is trained using a plurality of forces (605) having different normal force components.

25. The method according to any one of claims 14 to 24, wherein the multiple forces (605) applied by the indenter (600) are measured by a force sensor (610) located within the indenter (600) or adjacent to the indenter (600).

26. Each simulated force vector (F) has a normal force component (F Ｎ a) and the first shear force component (F Ｓ１ a) and the second shear force component (F Ｓ２ The method according to any one of claims 14 to 25, comprising a)

27. The method according to claim 1, wherein the feedforward neural network (FFNN) directly maps a barometric pressure sensor (400) to a force map (FM).

28. Before force inference, the aforementioned feedforward neural network (FFNN) performs the following steps: - A step (T3_1) in which a plurality of force tests are performed on the sensor device (10), Each force test comprises step (T3_1), which includes applying a force (605) with a single indenter (600) at a position on the measurement surface (210) of the sensor device (10), simultaneously measuring the force (605) applied by the indenter (600), and simultaneously measuring the pressure value (R) with a pressure sensor (400). - Step (T3_2) for each force test, in which a corresponding simulation is performed using the finite element model (10a) of the sensor device (10), Each simulation includes applying a simulated force (605a) to the simulated measurement surface (210a) of the finite element model (10a), thereby calculating a simulated force map (FMa) on the simulated measurement surface (210a). The simulated force map (FMa) includes a plurality of simulated force vectors (Fa), Step (T3_2) is performed such that the simulated force (605a) corresponds to the measured force (605) and is applied to a position on the simulated measurement surface (210a) that corresponds to the position on the measurement surface (210), - A step (T3_3) of training the feedforward neural network (FFNN) using the measured pressure value (R) and the corresponding calculated simulated force map (FMa), A method according to any one of claims 1 to 27, trained in [a specific manner].

29. The method according to claim 28, wherein the force test for training the feedforward neural network (FFNN) is performed using a plurality of indenters (600), each indenter having its own indenter shape.

30. The method according to claim 29, wherein the indenter shape is selected from the group comprising at least a tip, a circular shape, a triangular cross-section, a square cross-section, a hemisphere, a cube, and a cylinder.

31. The method according to claim 29 or 30, wherein the simulation is performed with a simulated force (605a) based on a simulated indenter (600a) having a simulated indenter shape corresponding to the actual indenter shape used in the corresponding force test.

32. The method according to any one of claims 28 to 31, wherein the feedforward neural network (FFNN) is trained using a plurality of different indenter shapes.

33. The method according to any one of claims 28 to 32, wherein the feedforward neural network (FFNN) is trained using a plurality of indenters (600) of different sizes.

34. The method according to any one of claims 28 to 33, wherein the feedforward neural network (FFNN) is trained by indenters (600) to which each shear force is applied for at least a portion of the force tests to train the feedforward neural network (FFNN).

35. The method according to any one of claims 28 to 34, wherein each measured force (605) includes a normal force component, a first shear force component, and a second shear force component.

36. In the measured force (605), the first shear force component corresponds to the first shear force, and the second shear force component corresponds to the second shear force. The method according to claim 35, wherein the first shear force is perpendicular to the second shear force.

37. The method according to claim 35 or 36, wherein the feedforward neural network (FFNN) is trained using a plurality of forces (605) having different shear force components.

38. The method according to any one of claims 35 to 37, wherein the feedforward neural network (FFNN) is trained using a plurality of forces (605) having different normal force components.

39. The method according to any one of claims 28 to 38, wherein multiple forces (605) are measured by force sensors (610) located within or next to the indenter (600).

40. Each simulated force vector (Fa) is the vertical resistance component (F Ｎ ), the first shear force component (F Ｓ１ a), and the second shear force component (F Ｓ２ a), and the method according to any one of claims 28 to 39.

41. The method according to any one of claims 1 to 40, wherein pressure values ​​(R) based on a calculated force map (FM) are read out simultaneously or for a predetermined period of time.

42. The force map (FM) is mm ２ A force vector (F) of at least 0.25 per unit, mm ２ A force vector (F) of at least 0.5 per unit, mm ２ A force vector (F) of at least 0.75 per unit, mm ２ A force vector (F) of at least 1 per unit, mm ２ A force vector (F) of at least 1.5 per unit, or mm ２ It must contain at least two force vectors (F), The force map (FM) is mm ２ Maximum force vector (F) of 0.25 per unit, mm ２ Maximum force vector (F) of 0.5 per unit, mm ２ Maximum force vector (F) of 0.75 per unit, mm ２ Maximum force vector (F) of 1, mm ２ A maximum force vector (F) of 1.5 per unit, or mm ２ The force vector (F) at the point of contact contains a maximum of 2. The method according to any one of claims 1 to 41, wherein at least one of the above.

43. Each force vector (F) has a normal force component (F Ｎ ) and the first shear force component (F Ｓ１ ) and the second shear force component (F Ｓ２ The method according to any one of claims 1 to 42, comprising:

44. First shear force component (F Ｓ１ ) corresponds to the first shear force, and the second shear force component (F Ｓ２ ) is designed to withstand the second shear force, The method according to claim 43, wherein the first shear force is perpendicular to the second shear force.

45. The method according to any one of claims 1 to 44, further comprising reading a temperature value from the pressure sensor (400) and providing temperature information or a temperature map of the sensor device (10) based on the temperature value.

46. A method (T2) for training a reconstructed network (RN), The reconstruction network (RN) maps the virtual sensor (400a) of the finite element model (10a) of the sensor device (10) to a force map (FM), The sensor device (10) includes a plurality of pressure sensors (400), The device comprises a pressure sensor (400) and a follow-up layer (200) that covers the pressure sensor (400) and provides a measurement surface (210), The force map (FM) includes a plurality of force vectors (F), - Each virtual sensor (400a) is provided with one or more virtual sensor points (410a), and each point has a value (S), - The reconstruction network follows these steps: - A step (T2_1) in which a plurality of simulations are performed in the finite element model (10a), Each simulation comprises simultaneously applying one or more simulated forces (605a) to the simulated measurement surface (210a) of the finite element model (10a), thereby calculating a simulated force map (FMa) on the simulated measurement surface (210a). The simulated force map (FMa) includes a plurality of simulated force vectors (Fa), Step (T2_1) involves calculating the value (S) of the corresponding virtual sensor point using the finite element model (10a), - A step (T2_2) of training the reconstructed network (RN) using the calculated simulated force map (FMa) and the corresponding calculated virtual sensor point values ​​(S), A method for training a reconstructed network (RN) (T2) which is trained using the method described above.

47. The method according to claim 46, wherein the simulated force (605a) applied to the simulated measurement surface (210a) is generated based on a simulated indenter (600a) having a simulated indenter shape.

48. The method according to claim 47, wherein the simulated indenter shape is selected from the group comprising at least a tip, a circle, a triangular cross-section, a square cross-section, a hemisphere, a cube, and a cylinder.

49. The method according to any one of claims 46 to 48, wherein the reconstruction network (RN) is trained using a plurality of different simulated indenter shapes.

50. The method according to any one of claims 46 to 49, wherein the reconstruction network (RN) is trained using a plurality of sized dummy indenters (600a).

51. The method according to any one of claims 46 to 50, wherein the reconstruction network (RN) is trained by at least a portion of a simulation comprising the simultaneous application of simulated forces (605a) generated based on two or more simulated indenters (600a).

52. The method according to any one of claims 46 to 51, wherein the reconstruction network (RN) is trained by at least a portion of a simulation which includes applying a simulated force (605a) generated based on only one simulated indenter (600a).

53. Each simulated force vector (F) has a normal force component (F Ｎａ ) and the first shear force component (F Ｓ１ a) and the second shear force component (F Ｓ２ The method according to any one of claims 46 to 52, including (a).

54. Among the simulated force vector (Fa), the first shear force component (F Ｓ１ a) corresponds to the first shear force, and the second shear force component (F Ｓ２ a) corresponds to the second shear force, The method according to claim 53, wherein the first shear force is perpendicular to the second shear force.

55. The method according to any one of claims 46 to 54, wherein the reconstructed network (RN) is trained using a plurality of simulated forces (605a) having different shear force components.

56. The method according to any one of claims 46 to 55, wherein the reconstruction network (RN) is trained using a plurality of simulated forces (605a) having different normal force components.

57. The method according to any one of claims 46 to 56, wherein the reconstructed network (RN) is used in the method according to claim 2 or in any one of the claims dependent on claim 2.

58. The force map (FM) is mm ２ A force vector (F) of at least 0.25 per unit, mm ２ A force vector (F) of at least 0.5 per unit, mm ２ A force vector (F) of at least 0.75 per unit, mm ２ A force vector (F) of at least 1 per unit, mm ２ A force vector (F) of at least 1.5 per unit, or mm ２ It must contain at least two force vectors (F), The force map (FM) is mm ２ Maximum force vector (F) of 0.25 per unit, mm ２ Maximum force vector (F) of 0.5 per unit, mm ２ Maximum force vector (F) of 0.75 per unit, mm ２ Maximum force vector (F) of 1, mm ２ A maximum force vector (F) of 1.5 per unit, or mm ２ The force vector (F) at the point of contact contains a maximum of 2. The method according to any one of claims 46 to 57, wherein at least one of the above.

59. Each force vector (F) has a normal force component (F Ｎ ) and the first shear force component (F Ｓ１ ) and the second shear force component (F Ｓ２ The method according to any one of claims 46 to 58, including the following:

60. The first shear force component (F Ｓ１ ) corresponds to the first shear force, and the second shear force component (F Ｓ２ ) is designed to withstand the second shear force, The method according to claim 59, wherein the first shear force is perpendicular to the second shear force.

61. A method for training a transfer network (TN), The transfer network (TN) maps the barometric pressure sensor (400) of the sensor device (10) to a plurality of virtual sensors (400a) of the finite element model (10a) of the sensor device (10). The sensor device (10) is Multiple barometric pressure sensors (400), A follow-up layer (200) covers the pressure sensor (400) and provides a measurement surface (210). Equipped with, - Each virtual sensor (400a) comprises one or more virtual sensor points (410a), and each point has a value (S), - The transfer network (TN) follows these steps: - A step (T1_1) in which a plurality of force tests are performed on the sensor device (10), Step (T1_1) comprises each force test comprising applying a force (605) by one indenter (600) at a position on the measurement surface (210) of the sensor device (10), simultaneously measuring the force (605) applied by the indenter (600), and simultaneously measuring the pressure value (R) with the barometric pressure sensor (400), - Step (T1_2) of performing a corresponding simulation using the finite element model (10a) for each force test, Each simulation includes applying a simulated force (605a) to the simulated measurement surface (210a) of the finite element model (10a), thereby calculating a simulated force map (FMa) on the simulated measurement surface (210a). The simulated force map (FMa) includes a plurality of simulated force vectors (Fa), Step (T1_2) is to apply the simulated force (605a) to a position on the simulated measuring surface (210a) that corresponds to the measured force (605) and the position on the measuring surface (210), The steps include: calculating the value (S) of the corresponding virtual sensor point using the finite element model (10a); - A step (T1_3) to train a transfer network using the measured pressure value (R) and the corresponding calculated virtual sensor point value (S), A method for training a transfer network (TN) that is trained using [a specific method / tool].

62. The method according to claim 61, wherein the force test for training the transfer network (TN) is performed using a plurality of indenters (600), each indenter having its own indenter shape.

63. The method according to claim 62, wherein the indenter shape is selected from the group comprising at least a tip, a circular shape, a triangular cross-section, a square cross-section, a hemisphere, a cube, and a cylinder.

64. The method according to claim 62 or 63, wherein the simulation is performed by a simulated force (605a) based on a simulated indenter (600a) having a simulated indenter shape corresponding to the actual indenter shape used in the corresponding force test.

65. The method according to any one of claims 61 to 64, wherein the transfer network (TN) is trained using a plurality of different indenter shapes.

66. The method according to any one of claims 61 to 65, wherein the transfer network (TN) is trained using a plurality of indenters (600) of different sizes.

67. The method according to any one of claims 61 to 66, wherein the transfer network (TN) is trained by indenters (600) to which each shear force is applied for at least a portion of the force test to train the transfer network (TN).

68. The method according to any one of claims 61 to 67, wherein each measured force (605) includes a normal force component, a first shear force component, and a second shear force component.

69. In the measured force (605), the first shear force component corresponds to the first shear force, and the second shear force component corresponds to the second shear force. The method according to claim 68, wherein the first shear force is perpendicular to the second shear force.

70. The method according to claim 68 or 69, wherein the transfer network (TN) is trained using a plurality of forces (605) having different shear force components.

71. The method according to any one of claims 68 to 70, wherein the transition network (TN) is trained using a plurality of forces (605) having different normal force components.

72. The method according to any one of claims 61 to 71, wherein the force (605) applied by the indenter (600) is measured by a force sensor (610) located inside the indenter (600) or adjacent to the indenter (600).

73. Each simulated force vector (F) has a normal force component (F Ｎ ) and the first shear force component (F Ｓ１ ) and the second shear force component (F Ｓ２ The method according to any one of claims 61 to 72, comprising:

74. The method according to any one of claims 61 to 73, wherein the transfer network is used in the method according to claim 2 or in any one of the claims dependent on claim 2.

75. A method (T3) for training a feedforward neural network (FFNN), The aforementioned feedforward neural network (FFNN) calculates a force map (FM) on the measurement surface (210) of the sensor device (10) based on the pressure value (R) of the barometric pressure sensor (400). The sensor device (10) is Multiple barometric pressure sensors (400), The device comprises a pressure sensor (400) covered by a follow-up layer (200) that provides a measurement surface (210), The force map (FM) includes a plurality of force vectors (F), The aforementioned feedforward neural network (FFNN) is constructed in the following steps: - A step (T3_1) in which a plurality of force tests are performed on the sensor device (10), Each force test comprises step (T3_1), which includes applying a force with one indenter (600) at a position on the measurement surface (210) of the sensor device (10), simultaneously measuring the force (605) applied by the indenter (600), and simultaneously measuring the pressure value (R) with a pressure sensor (400), - Step (T3_2) for each force test, in which a corresponding simulation is performed using the finite element model (10a) of the sensor device (10), Each simulation includes applying a simulated force (605a) to the simulated measurement surface (210a) of the finite element model (10a), thereby calculating a simulated force map (FMa) on the simulated measurement surface (210a). The simulated force map (FMa) includes a plurality of simulated force vectors (Fa), Step (T3_2) is performed, wherein the simulated force corresponds to the measured force (605) and is applied to a position on the simulated measurement surface (210a) that corresponds to the position on the measurement surface (210), - A step (T3_3) of training the feedforward neural network (FFNN) using the measured pressure value (R) and the corresponding calculated simulated force map (FMa), A method for training a feedforward neural network (FFNN) (T3), which is trained using the following method.

76. The method according to claim 75, wherein the force test for training the feedforward neural network (FFNN) is performed using a plurality of indenters (600), each indenter having its own indenter shape.

77. The method according to claim 76, wherein the indenter shape is selected from the group comprising at least a tip, a circular shape, a triangular cross-section, a square cross-section, a hemisphere, a cube, and a cylinder.

78. The simulation is performed by a simulated force (605a) based on a simulated indenter (600a) having a simulated indenter shape corresponding to the actual indenter shape used in the corresponding force test, according to the method of claim 76 or 77.

79. The method according to any one of claims 75 to 78, wherein the feedforward neural network (FFNN) is trained using a plurality of different indenter shapes.

80. The method according to any one of claims 75 to 79, wherein the feedforward neural network (FFNN) is trained using a plurality of indenters (600) of different sizes.

81. The method according to any one of claims 75 to 80, wherein the feedforward neural network (FFNN) is trained by the indenter (600) to which each shear force is applied for at least a portion of the force tests to train the feedforward neural network (FFNN).

82. The method according to any one of claims 75 to 81, wherein each measured force (605) includes a normal force component, a first shear force component, and a second shear force component.

83. In the measured force (605), the first shear force component corresponds to the first shear force, and the second shear force component corresponds to the second shear force. The method according to claim 82, wherein the first shear force is perpendicular to the second shear force.

84. The method according to claim 82 or 83, wherein the feedforward neural network is trained using a plurality of forces (605) having different shear force components.

85. The method according to any one of claims 82 to 84, wherein the feedforward neural network is trained using a plurality of forces (605) having different normal force components.

86. The method according to any one of claims 75 to 85, wherein multiple forces (605) are measured by force sensors (610) located within or next to the indenter (600).

87. Each simulated force vector (F) has a normal force component (F Ｎ a) and the first shear force component (F Ｓ１ a) and the second shear force component (F Ｓ２ The method according to any one of claims 75 to 86, comprising a)

88. Among the simulated force vector (Fa), the first shear force component (F Ｓ１ a) corresponds to the first shear force, and the second shear force component (F Ｓ２ a) corresponds to the second shear force, and The method according to claim 87, wherein the first shear force is perpendicular to the second shear force.

89. The method according to any one of claims 75 to 88, wherein the feedforward neural network is used in the method according to claim 1 or the method according to any one claim dependent on claim 1.

90. The force map (FM) is mm ２ A force vector (F) of at least 0.25 per unit, mm ２ A force vector (F) of at least 0.5 per unit, mm ２ A force vector (F) of at least 0.75 per unit, mm ２ A force vector (F) of at least 1 per unit, mm ２ A force vector (F) of at least 1.5 per unit, or mm ２ It must contain at least two force vectors (F), The force map (FM) is mm ２ Maximum force vector (F) of 0.25 per unit, mm ２ Maximum force vector (F) of 0.5 per unit, mm ２ Maximum force vector (F) of 0.75 per unit, mm ２ Maximum force vector (F) of 1, mm ２ A maximum force vector (F) of 1.5 per unit, or mm ２ The force vector (F) at the point of contact contains a maximum of 2. The method according to any one of claims 75 to 89, wherein at least one of the above.

91. Each force vector (F) has a normal force component (F Ｎ ) and the first shear force component (F Ｓ１ ) and the second shear force component (F Ｓ２ The method according to any one of claims 75 to 90, comprising:

92. The first shear force component (F Ｓ１ ) corresponds to the first shear force, and the second shear force component (F Ｓ２ The method according to claim 91, wherein the first shear force corresponds to a second shear force, and the first shear force is perpendicular to the second shear force.

93. In the method according to any one of claims 1 to 92, the sensor device (10) is a force-sensing sensor device (10), Flexible circuit board (300) and Multiple pressure sensors (400) attached to the flexible circuit board (300), A rigid core (100) to which the flexible circuit board (300) surrounds and to which the flexible circuit board (300) is attached, wherein the rigid core (100) is at least partially covered by the flexible circuit board (300) together with the pressure sensor (400) which protrudes away from the rigid core (100), A follow-up layer (200) that covers the pressure sensor (400) and provides a measurement surface (210), Equipped with, The method according to any one of claims 1 to 92.

94. The method according to claim 93, wherein the rigid core (100) is dome-shaped.

95. The method according to claim 93 or 94, wherein the rigid core (100) comprises a plurality of planar portions, and each of the pressure sensors (400) is arranged on one of the plurality of planar portions.

96. The method according to any one of claims 93 to 95, wherein the follow-up layer (200) contains or is made of a plastic material or rubber.

97. The method according to claim 96, wherein the plastic material is thermoplastic, elastomer, thermoplastic elastomer, or thermosetting.

98. The method according to any one of claims 93 to 97, wherein the tracking layer (200) relays the force applied to the measurement surface (210) to at least a portion of the pressure sensor (400).

99. The method according to any one of claims 93 to 98, wherein a plurality of pressure sensors (400) are connected by conductive paths on the flexible circuit board (300).

100. The method according to any one of claims 93 to 99, wherein the flexible circuit board (300) is shaped like an asterisk.

101. The method according to any one of claims 93 to 100, wherein the flexible circuit board (300) comprises a plurality of arms (310, 320, 330, 340, 350, 360) connected at a central portion (305).

102. The method according to any one of claims 93 to 101, wherein the plurality of pressure sensors (400) are arranged at least in one of the following arrangements: at a distance of at least 1 mm, at least 2 mm, at least 3 mm, at least 4 mm, or at least 5 mm, and at a distance of up to 1 mm, up to 2 mm, up to 3 mm, up to 4 mm, or up to 5 mm.

103. The method according to any one of claims 93 to 102, wherein the sensor device (10) is at least one of the tip of the robot and the operating element of the robot.

104. The method according to any one of claims 93 to 103, wherein the rigid core (100) is a part manufactured by three-dimensional fabrication.

105. A force inference module for a force-sensing sensor device (10), wherein the force inference module is configured to perform the method described in any one of claims 1 to 104.

106. A sensor device (10) for detecting force, wherein the sensor device (10) Flexible circuit board (300) and Multiple pressure sensors (400) attached to the flexible circuit board (300), A rigid core (100) to which the flexible circuit board (300) is attached and which surrounds the flexible circuit board (300), wherein the flexible circuit board (300) at least partially covers the rigid core (100) with a pressure sensor (400) that protrudes away from the rigid core (100), A follow-up layer (200) that covers the pressure sensor (400) and provides a measurement surface (210), The force inference module according to claim 105, A force-detecting sensor device (10) is provided.

Images