Torque measuring device for a stress wave gear

Abstract

Torque measuring device (100) for a stress wave transmission, the torque measuring device comprising: - a plurality of strain sensors (10) coupled to a flexspline of the stress wave transmission, wherein the strain sensors (10) are designed to detect stretching and / or compressive deformations and to provide them as measurement signals; - a readout circuit (20) which is coupled to the plurality of strain sensors (10) and which is designed to preprocess the measurement signals and to transmit them to a computer device (30); and - the computer device (30), which comprises a neural network and is designed to calculate a torque acting on the flexspline from the preprocessed measurement signals using machine learning; characterized in that the neural network is designed as a multi-layer network and as a recurrent network.

Description

[0001] The present invention relates to a torque measuring device for a stress wave transmission.

[0002] An expansion wave gear, also known as a strain wave gear, or sliding wedge gear, is a gear with an elastic transmission element characterized by high gear ratios and rigidity. A strain wave gear comprises three elements: an elliptical steel disc with a rolling bearing and a thin, deformable outer ring, a deformable cylindrical steel bushing with external gearing, the so-called flexspline, and a rigid cylindrical outer ring with internal gearing, the circular spline. In other words, a non-circular cam of the wave generator is part of the drive shaft and also the inner ring of the wave generator bearing.

[0003] The use of the flexibility of the flexspline element in a strain wave drive to measure torque indirectly via the strains in this part by applying strain gauges to the surface was first proposed by M. Hashimoto et al. ("A joint torque sensing technique for robots with harmonic drives", Proceeding of IEEE International Conference on Robotics and Automation, Vol. 2, pp. 1034-1039, April 1991).

[0004] Multiple expansion shaft gears, also known as harmonic drive gears, are widely used in the industrial, aerospace, and automotive industries. Due to their compactness while transmitting high torques at high gear ratios, their virtually backlash-free design, and acceptable efficiency, these gears are also very popular in robotic applications such as industrial manipulator arms.

[0005] They usually consist of a circular ring gear with internal teeth and a number of teeth N, a flexible element with external teeth and a number of teeth N-2, in a typical configuration, and a wave generator, which is a non-circular bearing, typically elliptical.

[0006] In some applications, and especially in robotics, it is necessary to measure the torque applied to the environment for safety and control purposes.

[0007] A common solution for torque measurement is 1.) sensing the current of the motor driving the gearbox, which is inexpensive but inaccurate, or 2.) adding a separate so-called flexure structure with attached strain sensors to the output side of the gearbox, which requires additional mass and takes up space but allows for a more accurate measurement.

[0008] One concept that can accurately measure the torque on the output side of the gear without additional weight and space is the application of a strain sensor directly to flexible elements of the gear.

[0009] However, existing solutions suffer from several problems, such as insufficient long-term stability and reduced precision.

[0010] The purpose of the present invention is to provide a strain wave transmission with an integrated torque sensor that does not suffer from such disadvantages.

[0011] Several authors propose variations of this approach using different sensor orientations and signal processing techniques to prevent disturbances, such as Tagirad et al., "Intelligent Built-In Torque Sensor for Harmonic Drive Systems," IEEE Instrumentation and Measurement Technology Conference, May 1997, which proposed eliminating high-frequency torque ripples from gear mesh vibrations and low-frequency disturbances caused by axis misalignment by applying a Kalman filter.

[0012] There are several published patents describing variations and details of this technology.

[0013] The documents JP 2000 320 622 A and DE 10 2004 041 394 A1 describe the permanent connection of wires to strain gauges attached to the flexible element of a strain wave gear.

[0014] The documents US 6,170,340 B1 and JP 2016 45055 A describe special arrangements of the strain gauges, which presumably have advantages in terms of minimizing disturbances.

[0015] The document US 2004 / 0079174 A1 describes the arrangement of the strain gauges and wires in a common unit that is collectively attached to the flexible element of the gear.

[0016] The documents US 6,840,118 B2 and CN 105 698 992 A deal with methods for suppressing unwanted disturbances in the torque measurement signal, such as torque ripples.

[0017] The documents RU 2 615 719 C1 and WO 2010 / 142318 A1 describe other methods for measuring torque in strain wave gears, such as applying strain gauges to the circular gear ring.

[0018] DE 103 42 479 A1 describes a torque detection device for a shaft gear with a strain gauge unit having a strain gauge pattern. This pattern includes circular arc-shaped detection segments A and B and three connection areas for external wiring, one of which is formed between the detection segments and the others are formed at the opposite ends thereof.

[0019] DE 102 17 020 A1 describes a method for determining and compensating periodically occurring disturbance torques in a harmonic drive transmission that is arranged downstream of a drive motor.

[0020] DE 103 21 210 A1 describes a method for measuring the transmitted torque in a shaft gear, wherein several sets of strain sensors are mounted on the surface of the flexible external gear, and the method comprises the following steps: amplifying and adjusting the output signals of each set such that the rotational ripple components contained in the output signals, which are caused by a distortion of the flexible external gear and are not related to the transmitted torque, are eliminated or controlled,

[0021] Combining the output signals of the sets and outputting a measurement signal in which the rotational ripple component can be fully compensated by a suitable gain adjustment of the output signals.

[0022] The object of the present invention is to provide a strain wave transmission with an integrated torque sensor which does not suffer from the above-mentioned disadvantages. Summary of the invention

[0023] According to the invention, this object is achieved by the torque measuring device for a stress wave transmission according to claim 1 and by a corresponding stress wave transmission according to claim 9. Appropriate further developments can be found in the dependent claims.

[0024] According to a first aspect of the present invention, a torque measuring device according to the invention for a stress wave transmission is provided, in which it is provided that the torque measuring device comprises: a plurality of strain sensors coupled to a flexspline of the stress wave transmission, wherein the strain sensors are designed to detect stretching and / or compressive deformations and to provide them as measurement signals.

[0025] Furthermore, the torque measuring device according to the invention comprises a readout circuit which is coupled to the plurality of strain sensors and which is designed to preprocess the measurement signals and to transmit them to a computer device.

[0026] Furthermore, the torque measuring device according to the invention comprises a computer device which comprises a neural network and which is designed to calculate a torque acting on the flexspline from the preprocessed measurement signals using machine learning.

[0027] Further developments can be found in further exemplary embodiments. Measuring the strain on the flexspline of an expansion shaft gear can be used to estimate the torque applied to the gear's output.

[0028] However, due to the inherent principle of such gears, the strains measured on the flexspline do not only originate from the torsional loads around the rotation axis of the gear, which transmits the torque to be measured, but also from periodically varying bending moments, position-dependent stiffnesses and other parasitic effects.

[0029] Therefore, achieving high torque measurement precision / resolution is difficult. Various approaches exist to mitigate these undesirable effects, such as those in JW Sensinger and RF Wehr, "Improved Torque Fidelity in Harmonic Drive Sensors Through the Union of Two Existing Strategies," IEEE / ASME Transactions on Mechatronics, August 2006, which uses an intelligent array of strain gauges aligned by Wheatstone bridges, or Tagirad et al., "Intelligent Integrated Torque Sensor for Harmonic Drive Systems," IEEE Instrumentation and Measurement Technology Conference, May 1997, which applies a Kalman filter to reduce torque ripple effects.

[0030] However, all these approaches still suffer from relatively low accuracy compared to external torque sensor structures, and some result in signal delay, which is undesirable in real-time control applications such as robotics.

[0031] As the main parasitic effects, the torque ripple, oscillating at twice the output and input frequency of the gear, forms a spatial waveform that is almost constant in a coordinate system oriented to the gear's flexspline, but from other reference frames it may look like a temporal oscillation.

[0032] In other words, the present invention enables the spatial shape of the flexspline to be determined by attaching a plurality of strain sensors to the free surfaces of this part. The signals from the plurality of strain sensors are then preprocessed, such as amplified, normalized, or combined by Wheatstone bridges, and fed to a microprocessor, which applies a regression algorithm to extract the torque—or, in other words, the signal of interest.

[0033] To compensate for the influence of other physical quantities, it may also be useful to enter additional measurements such as temperature, angular velocity or position into the algorithm.

[0034] The present invention can be used for robot applications such as industrial manipulator arms or flexible manipulator arms.

[0035] In a preferred embodiment, a dense pattern of individual strain gauges is attached or coated to the cylindrical surface adjacent to the gear teeth on the flexible gear ring of a cup- or sleeve-shaped strain wave gear.

[0036] Another possible location is the lateral surface of the cup-shaped flexspline. This pattern would cover a large portion of the surface, the entire circumference, in one or more adjacent rows.

[0037] The strain gauges may all have an identical orientation, inclined at approximately +45 degrees with respect to a circumferential center of the cylindrical surface, or a varying orientation, varying within a range of approximately +45 degrees and -45 degrees or +75 degrees and -75 degrees.

[0038] The signals from the strain gauges are preprocessed and then fed into a multi-layered neural network as inputs, optionally together with other useful sensor signals.

[0039] The neural network is trained with data from experiments, for example, measuring the sensor outputs and a precise torque with an external high-precision sensor on a test bench in a large number of trials under different conditions.

[0040] The reference signal could also be post-processed to remove any remaining torque ripple before being used as a training or verification data signal.

[0041] The training of the neural network is usually performed offline using global optimization methods such as genetic algorithms or gradient methods such as backpropagation.

[0042] This advantageously allows the training of the neural network to correctly capture the deformation as a complex spatial function and to determine a corrected, more accurate torque accordingly.

[0043] However, it might also be possible to use adaptive online learning methods or reinforcement learning during operation under certain settings by the computer device.

[0044] The present invention advantageously enables the parasitic components in the sensor signals measured at the flexible ring gear of a strain wave gear to be eliminated or mitigated much more efficiently than conventional methods, which can increase the precision and sensitivity of torque measurement without introducing delay. The absence of algorithm-induced delays requires the use of a feed-forward network or other methods that do not depend on past values. For example, using a so-called "recurrent network" or a recurrent neural network.

[0045] Preferably, the computer device is designed to calculate the torque acting on the flexspline from the preprocessed measurement signals using a regression analysis.

[0046] According to the invention, the neural network is designed as a multi-layer network and as a recurrent network.

[0047] It is preferably provided that the readout circuit is designed to amplify and / or standardize the measurement signals.

[0048] Preferably, the readout circuit comprises a Wheatstone bridge.

[0049] Preferably, the computer device is further configured to calculate the torque acting on the flexspline based on at least one of the following parameters, in addition to the pre-processed measurement signals using machine learning: i) ambient temperature; ii) temperature of a component of the stress wave gear; iii) angular velocity of a component of the stress wave transmission; and iv) Position of a component of the stress wave gear.

[0050] Preferably, the strain sensors are designed as strain gauges, which are further configured to be arranged on the lateral surface of a cup-shaped flexspline. This can be achieved by gluing or coating.

[0051] Preferably, the method can be used without such additional measured values, i.e., the torque acting on the flexspline can be calculated directly. This offers the advantage of reduced complexity.

[0052] Preferably, the computer device is designed to train the neural network using test data from a stress wave transmission.

[0053] Further measures improving the invention are presented below together with the description and the figures of the drawings of the invention.

[0054] The accompanying drawings illustrate embodiments and, in conjunction with the description, serve to explain concepts of the present invention.

[0055] Other embodiments and many of the aforementioned advantages will become apparent with reference to the figures of the drawings. The elements illustrated in the figures of the drawings are not necessarily drawn to scale relative to one another. Short description of the characters

[0056] They show: Fig. 1: a schematic representation of a torque measuring device according to an embodiment of the present invention; and Fig. 2: a schematic representation of a torque measuring device according to an embodiment of the present invention. Detailed description of the implementation examples

[0057] In the figures of the drawings, the same reference symbols designate the same or functionally equivalent elements, parts, components or process steps, unless otherwise stated.

[0058] According to the Fig. 1 comprises the torque measuring device 100, a plurality of strain sensors 10 coupled to a flexspline of the stress wave transmission, a readout circuit 20 and a computer device 30.

[0059] According to the Fig. 2, the computing device 30 has a single-layer neural network. In the illustrated embodiment, the neural network is designed as a feedforward network.

[0060] Furthermore, the neural network can also be designed as a multi-layer network and provide the torque as the output value.

[0061] The strain gauges can be arranged as a dense pattern on the flexspline and evaluated like individual pixels of a planar torque distribution.

[0062] According to the Fig. 2, the computer device 30 enables an improved torque sensor system for expansion compensating gears and solves the problem that the sensor signal is not accurate enough or can only be calculated with a delay.

[0063] The computer device 30 is designed, for example, to process signals from a plurality of strain gauges on the flexspline and, if applicable, also from other sensors by means of a machine learning model and to obtain a correct, sufficiently accurate torque based thereon.

[0064] Although the present invention has been described above using preferred embodiments, it is not limited thereto, but can be modified in a variety of ways. In particular, the present invention can be changed or modified in a variety of ways without deviating from the essence of the invention.

[0065] It should also be noted that “comprehensive” and “comprising” do not exclude other elements or steps, and “a” or “an” does not exclude a plurality.

[0066] Furthermore, it should be noted that features or steps described with reference to one of the above embodiments may also be used in combination with other features or steps of other embodiments described above. Reference symbols in the claims are not to be considered as limiting.

Claims

[1] Torque measuring device (100) for a stress wave transmission, the torque measuring device comprising: - a plurality of strain sensors (10) coupled to a flexspline of the stress wave transmission, wherein the strain sensors (10) are designed to detect stretching and / or compressive deformations and to provide them as measurement signals; - a readout circuit (20) which is coupled to the plurality of strain sensors (10) and which is designed to preprocess the measurement signals and to transmit them to a computer device (30); and - the computer device (30), which comprises a neural network and which is designed to calculate a torque acting on the flexspline from the preprocessed measurement signals using machine learning; characterized by that the neural network is designed as a multi-layer network and as a recurrent network. [2] Torque measuring device (100) according to claim 1, characterized by that the computer device (30) is designed to calculate the torque acting on the flexspline from the preprocessed measurement signals using a regression analysis. [3] Torque measuring device (100) according to claim 1 or 2, characterized by that the readout circuit (20) is designed to amplify and / or standardize the measurement signals. [4] Torque measuring device (100) according to one of claims 1 to 3, characterized by that the readout circuit (20) comprises a Wheatstone bridge. [5] Torque measuring device (100) according to one of claims 1 to 4, characterized by that the computer device (30) is further designed to calculate the torque acting on the flexspline based on at least one of the following parameters in addition to the preprocessed measurement signals using machine learning: v) ambient temperature; (vi) temperature of a component of the stress wave gear; vii) angular velocity of a component of the stress wave transmission; and viii) Position of a stress wave gear component. [6] Torque measuring device (100) according to one of claims 1 to 5, characterized by that the strain sensors (10) are designed as strain gauges, which are further designed to be arranged on a lateral surface of a cup-shaped or sleeve-shaped flexspline. [7] Torque measuring device (100) according to one of claims 1 to 6, characterized by that the computer device (30) is designed to train the neural network using test data from a stress wave transmission. [8] Stress wave transmission comprising a torque measuring device (100) according to one of claims 1 to 7. [9] Stress wave transmission according to claim 8, further comprising a cup-shaped or sleeve-shaped flexspline, wherein the strain sensors (10) of the torque measuring device are designed as strain gauges which are arranged on the lateral surface of the cup-shaped or sleeve-shaped flexspline in a constant orientation with respect to a circumferential center of the lateral surface of the cup-shaped flexspline or are arranged in a varying orientation with respect to a circumferential center of the lateral surface of the cup-shaped or sleeve-shaped flexspline.

Images