A method of monitoring a condition of a gear cutting machine having a plurality of machine axes, a method for creating a training data set of a machine learning algorithm, a method of training a machine learning algorithm, and a device for monitoring a condition of a gear cutting machine having a plurality of machine axes

Abstract

This invention provides a method for monitoring the state of a machine tool (1) having a plurality of machine axes. In this method, at least a portion of the machine axes are systematically actuated during a test cycle, and associated state data is obtained by measurement. Based on this, EOL data related to the noise behavior of a gear system comprising a workpiece machined by a gear cutting machine is predicted. The reverse prediction of state data from EOL data is also disclosed.

Description

Technical Field

[0001] This invention relates to a method for monitoring the status of a gear cutting machine used for processing toothed workpieces, and more specifically to a gear cutting machine. Prior Technology

[0002] During the hard machining of toothed workpieces, specifically external or internal gears, in a gear cutting machine, manufacturing deviations naturally occur, manifesting as the discrepancy between the actual geometry of the workpiece and its specified nominal geometry. Among other things, manufacturing deviations can be caused by malfunctions or wear of various components of the gear cutting machine, or by improper assembly of components. For example, manufacturing deviations can be caused by drive components that move the gear cutting machine's slides to positions other than those specified by the machine controller, by worn bearings on the spindle, or by improperly connected machine parts resulting in insufficient vibration reduction.

[0003] Hard machining is typically the final step in workpiece machining. Following this step, the workpiece is assembled into a gear train. After assembly, the gear train undergoes a final test ("EOL test") on an EOL (End of Line) test bench. Specifically, this tests the noise behavior of the gear train. Manufacturing deviations in the workpiece often lead to unwanted noise excitation. Even the smallest manufacturing deviations that can occur during machining on a fault-free machine can cause noise. Especially in electrically driven applications, gear train components are sometimes the primary source of noise, and noise within the gear train is particularly disruptive.

[0004] It is desirable to predict the noise behavior caused by the workpiece before it is installed in the gear train in order to avoid costly disassembly. This becomes necessary if the gear train is found to be susceptible to noise during EOL testing. It is also necessary to be able to draw direct conclusions about specific causes in the gear cutting machine from the measured noise behavior of the gear train.

[0005] US20140256223A1 discloses a method for hard machining the tooth flanks of a gear cutting machine by correction and / or modification, wherein a gear pair or test device that engages with each other during transport is machined taking into account the individual mating sides, and wherein the tooth flanks of the relevant workpieces have periodic ripple correction or modification. The degree of rotational error is determined by means of a gear measuring device and / or by measuring the rotational distance error of the gear pair during transport. This measurement result serves as input values ​​to define the amplitude, frequency, and phase position of the periodic side ripple correction on the tooth flanks of the gear pair produced in the gear cutting machine. Summary of the Invention

[0006] [First-State Sample: Prediction Based on Noise Excitation]

[0007] In the first embodiment, the object of the present invention is to provide a method for monitoring the state of a gear cutting machine to predict early on the development of interference noise in a gear train containing workpieces processed by the gear cutting machine.

[0008] This objective is achieved by the method described in claim 1 of the patent application. Other specific examples are given in the subsidiary patent applications.

[0009] Therefore, a method for monitoring the state of a gear cutting machine having a plurality of machine axes is disclosed, comprising the following steps: a) Conduct a test cycle, wherein during the test cycle, at least a portion of the machine axis is systematically actuated and associated machine measurement data is obtained; b) Perform spectral analysis on the machine measurement data, including calculating the machine spectrum data from the machine measurement data; and c) Determine the predicted EOL spectrum based on machine spectrum data, wherein the predicted EOL spectrum data indicates the order in the expected excitation in the EOL spectrum when a workpiece machined by a gear cutter is mounted in a gear train and rolls on the mating gears in the gear train.

[0010] The method may include: d) Output predicted EOL spectral data or at least one quantity derived therefrom.

[0011] Within the context of the proposed method, machine measurement data is determined during a test cycle when the machine axis is selectively actuated. The test cycle occurs during a pause in the gear cutting machine's machining process; that is, during the test cycle, the gear cutting machine's machining tool is not engaged with the workpiece. Specifically, the machine measurement data may be acceleration values ​​determined by an accelerometer, position values ​​determined by a position sensor, and / or current values ​​determined by a current sensor. Spectral analysis (specifically, order analysis) is used to calculate machine spectral data (machine spectral state data) from the machine measurement data to determine the frequencies or orders at which the periodic excitation of machine components occurs when the machine axis is actuated.

[0012] Based on machine spectrum data, EOL spectrum data is predicted. This prediction is based on the following considerations: excitations from machine components propagate to the manufactured workpiece during machining. When the machining tool is dressed using the machine axis, these excitations also propagate to the dressed tool and further from the tool to the workpiece during machining. Using known kinematic links between individual machine components and considering selected machining parameters, it is possible to calculate how these excitations affect the manufactured workpiece. Specifically, such excitations can cause periodic deviations (ripples) on the tooth surface of the workpiece. The terms "kinematic link" or "kinematic chain" are used to describe how the movement of one component propagates to another component during the machining process. In the machining process, the kinematic link between the tool spindle and the workpiece spindle plays a central role, that is, how the movement of the tool is related to the movement of the workpiece. If the gear cutting machine is a machine used to machine a workpiece using production processes (such as production grinding or gear cutting machines), then this link is characterized by the rolling engagement between the workpiece and the tool and determined by the workpiece geometry and tool geometry. The periodic excitation of the tool, propagating to the side of the workpiece via this link as a ripple, can be readily calculated. Specifically, if the order of the excitation causing the tool's rotation relative to the tool is known, it is readily possible to calculate the order of the excitation generated by the ripples relative to the workpiece in the gear train. The ratio between these two orders will be referred to below as the "propagation factor". For all other components between propagating motions, it is also possible to calculate how the vibration of one of these components affects the other, and specifically, to calculate the corresponding propagation factor between the perturbation order of these components and the perturbation order obtained in the EOL spectrum.

[0013] Therefore, calculating and predicting EOL spectrum data may involve applying a propagation factor to the machine spectrum data, where the propagation factor depends on the motion linkage between the machine axis and the workpiece used to determine the machine spectrum data.

[0014] Therefore, predicted EOL spectrum data is obtained by a specified method. This predicted EOL spectrum data is based on testing in a real machine condition and therefore includes disturbance sources within the machine that may not be known a priori and therefore cannot be included in the calculation or simulation of the EOL spectrum data based purely on the known design of the machine. Thus, the method allows for reliable prediction of which orders of noise excitation will occur in the finished gear train. In this way, noise excitation can be predicted even before the workpiece is actually installed in the gear train. Affected workpieces can be rejected and, if necessary, inspected in more detail, thus avoiding the time-consuming disassembly of a fully assembled gear train. Furthermore, measures can be taken to identify and eliminate sources of error that lead to the expected development of noise.

[0015] It is worth noting that the determination of predicted EOL spectrum data is not necessarily a quantitative prediction of the noise intensity at various levels, but rather a qualitative indication of which levels are primarily "perturbation levels," that is, levels expected to have significant noise intensity. Specifically, predicted EOL spectrum data may include perturbation levels and associated intensity indicators, where the intensity indicators represent (possibly only very coarsely) an estimate of the expected perturbation intensity at the perturbation levels.

[0016] Specifically, EOL spectrum data can be predicted individually based on the actuated machine axis. That is, individual machine measurement data is determined for each machine axis activated during the test cycle, and individual machine spectrum data is calculated from these individual machine measurement data. Based on this calculation, individual EOL spectrum data is predicted according to the actuated machine axis. In this way, it is possible to predict which machine axis will cause which disturbance steps in the EOL spectrum data.

[0017] Preferably, steps a) through c) are repeated several times, wherein the workpiece is machined with a gear cutter between test cycles, and these test cycles are performed during machining pauses when the machining tool is not engaged with the workpiece. The evolution of the predicted EOL spectrum data as a function of the test cycle or time is then visualized and / or analyzed. This is based on the consideration that, based on a single test cycle, the predicted EOL spectrum data may sometimes have very few values. However, during machining, wear or failure of gear cutter components may occur, which then manifest themselves in significant changes to the predicted EOL spectrum data. Therefore, it is proposed to consider the temporal evolution of the predicted EOL spectrum data. By appropriately visualizing the temporal changes in the predicted spectrum data, perturbation orders expected to undergo significant changes in noise-excitation behavior over time can be visually easily identified. The evolution of the calculated EOL spectrum data can also be analyzed numerically. For example, numerical analysis may include regression analysis using appropriate regression functions (such as at least a second-order polynomial) on the expected perturbation intensity for selected or all orders of perturbation. For example, if the analysis shows that at least one perturbation order is expected to have a gradient in the perturbation strength that satisfies a certain warning criterion, then a warning indicator can be determined and output based on this.

[0018] If reference machine spectrum data is available for many reference machines to be evaluated in many different reference test cycles, reference EOL spectrum data can be predicted from each of these reference machine spectrum data. It is then possible to automatically evaluate the predicted EOL spectrum data of the machine to be evaluated by comparing these predicted EOL spectrum data with the predicted reference EOL spectrum data of the reference machines or quantities derived therefrom, and thus automatically inferring expected noise problems without any special knowledge or the need to measure the EOL spectrum as a basis for evaluation. Specifically, statistical analysis of the predicted reference EOL spectrum data can be performed for this purpose. For the underlying considerations of this method and other specific examples, reference can be made to Patent Application No. 111138213, entitled "Method of monitoring the condition of a machine tool," filed by the same applicant on the same date as this application, the contents of which are incorporated herein by reference in their entirety.

[0019] [Second-state pattern] [:] [Identify components causing noise issues] []

[0020] The reverse is also possible; that is, when the workpiece rolls on the mating gears in the gear train, the EOL measurement value on the EOL test bench is measured, and the spectrum analysis of the EOL measurement value is performed to determine the measurement spectrum data. From the EOL spectrum data, it can be concluded that the individual components of the gear cutting machine cause disturbance steps in the EOL spectrum data due to their state. The EOL measurement value can be determined by any suitable sensor on the EOL test bench, especially acceleration sensors and sensors used to determine rotational errors.

[0021] Therefore, a method for monitoring the state of a gear cutting machine having a plurality of machine axes is disclosed, comprising the following steps: a) Perform EOL testing on gear trains containing workpieces machined by a gear cutting machine, wherein in the EOL test, the workpieces in the gear train roll on the mating gears, and determine the relevant EOL measurement data; b) Perform spectral analysis on the EOL measurement data, including calculating the EOL spectral data from the EOL measurement data; and c) Determine the predicted state data based on EOL spectrum data, wherein the predicted state data of at least one machine axis indicates which order of that machine axis is consistent with the calculated EOL spectrum data.

[0022] The method may include: d) Output predicted state data or at least one quantity derived from it.

[0023] This method allows us to determine which machine axes and (if applicable) which components of these machine axes are responsible for disturbance noise that actually occurs after the workpiece being machined by the gear cutter is mounted in the gear train. If the motion linkages between the gear cutter components are known and the selected machining parameters are considered, this method also utilizes the fact that the order of disturbance noise related to the rotation of the workpiece in the gear train can be easily calculated from the order of measurement data determined by the gear cutter components.

[0024] Specifically, this method can be performed without the need for state measurement on the gear cutting machine itself. However, certain advantages arise when this method is combined with state measurement on the gear cutting machine. In this regard, the method may additionally include: e) Conduct a test cycle in which at least a portion of the machine axes is systematically actuated and associated machine measurement data is obtained; f) Perform spectrum analysis, including automatic machine measurement data and computer spectrum data; and g) Determine the predicted EOL spectrum data based on machine spectrum data, wherein the predicted EOL spectrum data indicates the expected excitation order in the EOL spectrum when a workpiece machined by a gear cutter is mounted in a gear train and rolls on mating gears in the gear train. The determination of the predicted state data includes comparing the EOL spectrum data calculated from the EOL measurement data with the predicted EOL spectrum data.

[0025] By comparing the EOL spectrum data determined by measurements during the test run of the gear train with the EOL spectrum data predicted from the measurement data of the gear cutting machine, the cause of disturbance noise can be determined with particular reliability.

[0026] [Third-State Sample: The Use of Machine Learning Algorithms] []

[0027] The methods discussed to date require knowledge of the kinematic linkages between the components of the gear cutting machine. In another embodiment, the present invention provides a method that makes it possible to predict the noise behavior of a gear train based on measurements of the machine's state, or to derive conclusions about the state of the gear cutting machine from the measured noise behavior of the gear train even without knowledge of the kinematic linkages. This method uses a trained machine learning algorithm whose input variables are the state data of the gear cutting machine and whose output variables are predicted EOL data, which represents the characteristics of the expected noise behavior of the gear cutting machine; or its input variables are EOL data and its output variables are predicted state data, which represents the characteristics of the expected state of the machine.

[0028] Use the following procedure to train machine learning algorithms: a) Conduct a test cycle in which at least a portion of the machine axis is systematically actuated and the associated status data is determined by measurement; b) When the gear cutter is in the state corresponding to the state data, at least one workpiece is processed by the gear cutter; c) Install the machined workpiece into the gear train; d) Perform an EOL test on the gear train, wherein the workpiece in the gear train rolls on the mating gears during the EOL test, and determine the associated EOL data; e) Store the state data and corresponding EOL data in the training dataset; f) Repeat steps a) to e) for multiple test cycles and processed workpieces, wherein the workpieces have the same nominal geometry and are processed under the same processing conditions; and g) Train machine learning algorithms using training datasets.

[0029] The training dataset therefore contains a large amount of state data and corresponding EOL data for multiple workpieces with the same nominal geometry, which are machined under the same machining conditions and installed in the same type of gear train.

[0030] Nominal geometry includes quantities such as the normal module of gear teeth, the number of teeth, and the helix angle, but may also include other quantities such as specified tooth vent modifications. Specifically, if the machine axis is moved in the same manner during machining operations, the machining state is considered identical. For example, if grinding is used as a machining process, if the workpiece is machined with the same radial feed, the same axial feed rate, and the same offset, if the tool rotational speed is the same for all workpieces, and if the grinding worm used has the same number of starts and the same pitch height for all workpieces, then the machining state is identical, resulting in the same rotational speed of the workpiece. If the grinding worm is a dressable grinding worm dressed by a dressing tool in the shape of a rotary disc, then the state during dressing is also part of the machining parameters, especially the spindle speed and the dressing tool speed during the dressing process.

[0031] The machine learning algorithm is trained using state data and corresponding end-of-life (EOL) data. Therefore, the machine learning algorithm can predict EOL data based on state data or vice versa, without needing to know the motion linkages between the components of the gear cutting machine.

[0032] Many different types of machine learning algorithms are known to be applicable in this context, and the structure of the training dataset can vary accordingly. Specifically, classification algorithms are suitable for practical implementation. For this purpose, the output variables can be reduced to a finite number of categories. For example, if the input variables are EOL data and the output variables are predicted state data, the predicted state data can consist of, for example, one real value per machine axis. Each value can then indicate, for example, the probability that the machine axis in question is responsible for the observed EOL data. The training data should then contain state data representing a single true state value for each machine axis and the associated EOL data. For example, if the input variables are state data and the output variables are predicted EOL data, then for a relatively small number of orders (especially important in practice), the predicted EOL data can consist of one real value per order. Each value can then indicate, for example, the predicted relative spectral intensity of the order in question. The training data should then contain the corresponding EOL data. Of course, completely different and even more complex outputs are also conceivable. For practical implementation, artificial neural networks (ANNs) or support vector machines (SVMs) are suitable. In a particularly simple example, the input variable can be state data, and the output variable can be a single real value that characterizes the global noise behavior of the entire gear train on the test bench. For example, random forests are suitable for machine learning algorithms that predict this value. With this value, problematic noise behavior can be easily detected, and measures can be taken to prevent affected components from being installed in the gear train.

[0033] Condition data can typically include various types of data relating to the state of a machine axis relative to its vibrational behavior. Specifically, condition data can include machine spectrum data as defined in the context of the first and second condition samples.

[0034] EOL data may also include various types of data related to the noise performance of the gear train. Specifically, EOL data may include EOL spectrum data as defined in the content background of the first and second states.

[0035] Training data can be stored in a database. The database can be located remotely from the monitoring machine. It can also be implemented in the cloud, for example, as a shared computer resource used by multiple users as a service. The evaluation computer can access the database to train machine learning algorithms. The evaluation computer is also preferably spatially separated from the machine tool. It is connected to the machine tool via a network connection. The evaluation computer does not need to be a single physical unit and can be implemented in the cloud.

[0036] The present invention further provides an apparatus for monitoring the state of a gear cutting machine having a plurality of machine axes, the apparatus comprising a processor and a storage medium thereon storing a computer program. When executed on the processor, the computer program causes at least a portion of the method steps of one of the methods explained above to be performed. The present invention further provides a corresponding computer program. The computer program may be stored on a non-volatile storage medium. Simple Explanation of the Diagram

[0037] Preferred embodiments of the invention are described below with reference to the accompanying drawings, which are for illustrative purposes only and not for limiting the scope of the invention. In the drawings, [Figure 1] shows a schematic diagram of the grinding machine; [Figure 2] A graph illustrating the interpretation of the spectral data obtained during the measurement period; [Figure 3] shows a sketch of an EOL test bench; [Figure 4A] and [Figure 4B] illustrate the extraction from a chart used to convert the steps of a machine component into the steps of a workpiece in a gear train mounted on an EOL test bench; [Figure 5] A graph illustrating the interpretation of spectral prediction data; [Figure 6] A schematic representation of the temporal evolution of the perturbation intensity at the perturbation step; [Figure 7] A sketch illustrating a machine learning algorithm; and [Figure 8] shows an excerpt of exemplary training data used for machine learning algorithms. Implementation

[0038] [Example structure of a grinding machine] []

[0039] Figure 1 illustrates a grinding machine 1 as an example of a gear cutting machine, which will also be abbreviated as "machine" below. Machine 1 has a machine bed 11, on which a tool carrier 12 is guided in a displaceable manner along the radial feed direction X. The tool carrier 12 carries an axial slide 13, which is guided in a displaceable manner relative to the tool carrier 12 along the axial feed direction Z. A grinding head 14 is mounted on the axial slide 13, which is rotatable about a rotation axis extending parallel to the X direction (the so-called A-axis) to accommodate the helix angle of the gear to be machined. The grinding head 14 also carries a displacement slide, on which a tool spindle 15 is displaceable relative to the grinding head 14 along the displacement direction Y. A worm-shaped grinding wheel (grinding worm) 16 is clamped on the tool spindle 15. The grinding worm 16 is driven by the tool spindle 15 to rotate about the tool axis B.

[0040] Machine tool 11 also supports a turret-shaped rotating workpiece carrier 20, which can rotate between at least three positions around the rotation axis C3. Two identical workpiece spindles are mounted diametrically opposite each other on the workpiece carrier 20, with only one workpiece spindle 21 having a corresponding custom tailstock 22 visible in Figure 1. Workpieces can be clamped on each of the workpiece spindles and driven to rotate around the workpiece axis C1 or C2. The workpiece spindle 21 visible in Figure 1 is in the machining position, in which the workpiece 23 clamped thereon can be machined using the grinding worm gear 16. The other workpiece spindle, offset by 180° and not visible in Figure 1, is in the workpiece changing position, from which the finished workpiece can be removed and a new blank can be clamped onto the workpiece. A dressing device 30 is mounted to the workpiece spindle at an offset of 90°.

[0041] Machine 1 therefore has a large number of movable components, such as sliders or spindles, which can move under the control of corresponding drives. These drives are often referred to in the technical world as "NC axes," "machine axes," or simply "axes." In some cases, this designation also includes components driven by drives, such as sliders or spindles.

[0042] Machine 1 also has a large number of sensors. As an example, only two sensors, 18 and 19, are schematically shown in Figure 1. Sensor 18 is an acceleration sensor (vibration sensor) for detecting vibrations of the housing of the grinding spindle 15. Sensor 19 is a position sensor for detecting the position of the axial slide 13 relative to the tool carrier 12 along the Z direction. However, in addition, machine 1 includes a plurality of other sensors. Specifically, these sensors include other position sensors for detecting the actual position of a linear axis in each case, rotation angle sensors for detecting the rotational position of a rotary axis in each case, current sensors for detecting the drive current of an axis in each case, and other vibration sensors for detecting vibrations of a drive assembly in each case.

[0043] All drive axes of machine 1 are digitally controlled by machine controller 40. Machine controller 40 includes several axis modules 41, a control computer 42, and a control panel 43. The control computer 42 receives operator commands and sensor signals from various sensors of machine 1 from the control panel 43, and calculates control commands for the axis modules 41 from these sensors. It also outputs operating parameters to the control panel 43 for display. Each axis module 41 provides a control signal for one of the machine axes at its output.

[0044] The monitoring device 44 is connected to the control computer 42. []

[0045] The monitoring device 44 may be a separate hardware unit associated with machine 1. It may be connected to control computer 42 via a known interface (e.g., via a known Ethernet standard) or via a network (e.g., via the Internet). It may be a spatial component of machine 1, or it may be spatially located away from machine 1.

[0046] During machine operation, the monitoring device 44 receives various measurement data from the control computer 42. Among the measurement data received by the control computer are sensor data directly acquired by the control computer 42 and data read by the control computer 42 from the axis module 41, such as data describing the target position of various machine axes and the target current consumption in the axis module.

[0047] The monitoring device 44 may have its own analog and / or digital sensor inputs to receive sensor data directly from other sensors as measurement data, depending on the situation. Other sensors are typically sensors that are not directly required to control the actual machining process, such as accelerometers used to detect vibration or temperature sensors.

[0048] The monitoring device 44 may alternatively be implemented as a software component of the machine controller 40, for example, running on the processor of the control computer 42, or may be implemented as a software component of the server 45, as described in more detail below. The server 45 has a processor 451 and a storage medium 452, which are indicated only schematically.

[0049] The monitoring device 44 communicates directly with the service server 45, or via the Internet and web server 47. The service server 45, in turn, communicates with the database server 46 and the database DB. These servers may be located remotely on machine 1. The server need not be a single physical entity. Specifically, the server may be implemented as a virtual unit in what is called the "cloud".

[0050] Server 45 communicates with terminal device 48 via web server 47. Specifically, terminal device 48 can run a web browser to visualize and evaluate the received data. Terminal device does not need to meet any specific computing power requirements. For example, terminal device can be a desktop computer, laptop computer, tablet computer, cellular phone, etc.

[0051] [Workpiece] [batch] [Processing]

[0052] The following describes how to process a workpiece using machine 1.

[0053] To process an unprocessed workpiece (blank), the workpiece is held by an automatic workpiece changer on the workpiece spindle in the workpiece changing position. The workpiece change occurs in parallel with the processing of another workpiece on a different workpiece spindle in the processing position. When the new workpiece to be processed is held and the processing of the other workpiece is completed, the workpiece carrier 20 rotates 180° around axis C3, causing the spindle with the new workpiece to be processed to move to the processing position. Before and / or during the rotary process, engagement is performed using associated engagement probes. For this purpose, the workpiece spindle 21 is set to rotate, and the position of the tooth clearance of the workpiece 23 is measured using engagement probe 24. The roll angle is determined based on this.

[0054] When the workpiece spindle carrying the workpiece 23 reaches the machining position, the workpiece 23 engages with the grinding worm 16 without collision by moving the tool carrier 12 along the X-axis. The workpiece 23 is then machined by rolling the grinding worm 16 into place. During machining, the workpiece advances continuously along the Z-axis at a constant radial X-feed. Additionally, the tool spindle 15 moves slowly and continuously along the offset axis Y to continuously utilize unused areas of the grinding worm 16 for machining (so-called offset movement).

[0055] The machining process is parallel to the workpiece, with the finished workpiece removed from the spindle of another workpiece and another blank clamped on this spindle.

[0056] If, after processing a certain number of workpieces, the grinding worm 16 has progressed to the point where it is too dull and / or its lateral geometry is too inaccurate, then the grinding worm is dressed. For this purpose, the workpiece carrier 20 is rotated ±90° so that the dressing device 30 is positioned relative to the grinding worm 16. The grinding worm 16 is then dressed using a dressing tool 33. This dressing tool is a rotary dressing wheel.

[0057] [Gear cutting machine testing cycle]

[0058] During machining pauses, the monitoring device 44 performs a test cycle that interacts with the machine controller 42 to check the status of individual or all components of machine 1. During this test cycle, selected portions or all machine axes are systematically actuated, and measurements are taken on the machine.

[0059] For example, each linearly displaceable component is displaced relative to its associated machine axis, and the instantaneous position of the component is continuously determined by the aforementioned position sensor. From there, the positional deviation between the specification (nominal position) and the measurement (actual position) is continuously determined, and this positional deviation is transmitted to the monitoring device 44. This operation can also be performed on a rotatably driven spindle, whereby a rotation angle sensor is then used to determine the positional deviation.

[0060] When the machine axis in question is activated, the vibration behavior of the selected machine axis is also determined. An acceleration sensor (vibration sensor) connected to these components is used for this purpose. The vibration measurement results are also transmitted to the monitoring device 44.

[0061] Furthermore, the power consumption of the drive motor of the machine axis is continuously determined when it is activated. A current sensor integrated in the axis module 41 can be used for this purpose, for example. The current measurement result is also transmitted to the monitoring device 44.

[0062] All of this can be done when a single machine axis is actuated individually. However, it is also possible to actuate two or more machine axes in a coupled manner, so that the machine's behavior is recorded when two or more machine axes are actuated simultaneously. In this case, for example, when a single machine axis is actuated, amplified vibrations greater than expected based solely on vibration behavior may occur, or controller errors that can only be determined when two machine axes are actuated synchronously may be detected.

[0063] [Status Information]

[0064] The monitoring device 44 determines various status data from the received measurement data. The status data allows for direct or indirect conclusions regarding the status of the machine or its individual components. Specifically, the status data includes spectral data obtained from the measurement data through spectral analysis. The complete spectrum or only the spectral intensity at selected discrete excitation frequencies can be determined.

[0065] Figure 2 shows an example of a spectrum obtained by filtering and FFT of the time signals of acceleration, position, or current sensors recorded during the actuation of a machine axis (here, the B-axis, i.e., the tool spindle). The spectrum in Figure 2 contains several clearly visible peaks at integer and non-integer multiples of the rotational frequency (order) of the machine axis in question.

[0066] For example, strong peaks at tool speeds and their multiples can indicate concentricity errors in the tool spindle. Peaks at higher tool speeds can indicate bearing damage in the tool spindle, and the bearing class can be inferred from the multiple. If the bearing class is known, it may be possible to identify the bearing causing the peak.

[0067] The monitoring device 44 transmits the obtained status data to the service server 45.

[0068] [EOL] [test]

[0069] Each finished workpiece is installed in the gear train. The gear train is tested on an EOL test bench before delivery. Refer to Figure 3 for a more detailed explanation of this situation.

[0070] Figure 3 shows machine 1 in a highly schematic manner, which has the various servers 45-47 and terminal devices 48 described above. An EOL test bench 2 is also shown in a highly schematic manner. The EOL test bench communicates with the server 45 via a web server 47.

[0071] As explained earlier, machine 1 has a plurality of sensors, including an acceleration sensor (vibration sensor) 51, a position sensor 52, and a current sensor 53. Also as previously explained, the machine uses these sensors to collect measurement data and sends the status data derived from them to the service server 45.

[0072] The EOL test bench also has various sensors, including accelerometers 54 and rotation angle sensors. The accelerometers 54 measure acoustic signals when a workpiece mounted in the gear train rolls on the mating gears. The EOL test bench then calculates resource data through spectrum analysis and sends it to the service server 45.

[0073] [Server]

[0074] Server 45 processes the received data, calculates other quantities from the received data as necessary, and stores the received data and calculated quantities in the database DB as required. Specifically, the server stores the following data: Status data with associated status identification data allows for the determination of the unique identification of the machine in the status data and the associated operations in the test cycle (specifically, the actuator axis of the machine), as well as the time of the test cycle; Process data for the machining process of each workpiece, and workpiece identification data that allows for unique identification of each workpiece; EOL data from the EOL test bench, and related workpiece identification data.

[0075] The service server can read and merge data from the database. Specifically, the service server can merge the EOL data of a specific workpiece with the associated process data and other machine state data that best represent the machine state as the processing state, and combine the data to form a dataset.

[0076] [Based on motion link pairs] [EOL] [Prediction of disturbances and noise in the spectrum]

[0077] The service server can perform qualitative predictions of the intensity of the perturbation order on the EOL test bench. For this purpose, the service server calculates the corresponding expected excitation spectrum (EOL spectrum) on the EOL test bench based on the spectrum in Figure 2.

[0078] In the calculation, the service server utilizes known motion linkages between the components of machine 1. Refer to Figures 4A and 4B for a more detailed explanation of this situation.

[0079] Figure 4A shows a partial example of the chart, where the known possible disturbance order of the B-axis (i.e., the tool spindle) and the corresponding expected disturbance order in the EOL spectrum are entered. In this example, these disturbance orders are at a fixed ratio of 3.45, which is given by the motion linkage between the tool spindle and the workpiece, i.e., by the rolling coupling between the tool and the workpiece, and determined by the geometry of the workpiece and the grinding worm. Symbolically, this ratio indicates how the vibration of the B-axis propagates as ripples on the tooth surface of the workpiece. This ratio can be calculated by considering the contact state between the grinding worm and the workpiece. It is referred to below as the "propagation factor". If the order of the components of the B-axis (e.g., bearing order and motor order) is known, the possible disturbance orders of the B-axis can be calculated in advance. The actual disturbance order of the B-axis can be determined by measurement.

[0080] Figure 4B shows a partial example of the graph, where the possible perturbation order of the Y-axis (i.e., the offset axis) and the corresponding expected perturbation order in the EOL spectrum are entered. On one hand, the graph distinguishes different components of the Y-axis (e.g., the ball screw drive BSD and drive motor), which can cause these perturbation orders; on the other hand, it distinguishes the perturbation orders in the EOL spectrum attributed to vibrations during workpiece machining (grinding) and dressing. Vibrations during grinding directly cause lateral ripples (waves) on the workpiece side. Vibrations during dressing initially cause lateral ripples on the grinding worm, and during grinding, these ripples are also converted into lateral ripples on the workpiece side. If the motion linkages and machining parameters during grinding and dressing are known, the corresponding propagation factor between the possible perturbation order of the Y-axis and the resulting perturbation order in the EOL spectrum can be easily calculated. The possible perturbation order of the Y-axis can be measured or calculated again.

[0081] This type of analysis can be performed on the possible perturbation order of the machine axis and the perturbation order obtained in the EOL spectrum for each machine axis involved in the grinding process.

[0082] Currently, the EOL spectrum is predicted based on the spectrum determined during the test cycle on the machine and the known propagation factor between the perturbation order of the machine axis and the associated perturbation order in the EOL spectrum. Refer to Figure 5 for a more detailed explanation of this situation. Figure 5 shows the expected predicted EOL spectrum when the B-axis test in the test cycle produces the spectrum of Figure 2. The predicted EOL spectrum is essentially the same as the spectrum in Figure 2, but extended along the horizontal axis by the propagation factor of 3.45 mentioned above as an example. However, the absolute signal values ​​in this EOL spectrum should be carefully examined: in short, for a given perturbation order, how strong the EOL signal will actually be depends not only on the strength of the corresponding perturbation order causing the machine axis, but also on a multitude of other factors during workpiece machining and the workpiece's mounting condition in the gear train. In this regard, the spectrum in Figure 5 does not allow for any quantitative statement regarding the expected signal strength. However, it also allows for the prediction of which perturbation levels are actually attributable to perturbation levels present in the EOL spectrum of the machine axis in question, and allows for qualitative estimation of the expected signal strength under these perturbation levels. For example, the spectrum in Figure 5 allows for a rough estimate of the signal strength under certain perturbation levels of interest, which cause perceived unwanted noise. These perturbation levels are circled as examples in Figure 5.

[0083] Overall, this allows for good prediction of the expected perturbation order with approximate signal strength, attributable to factors within the EOL.

[0084] For example, worn bearings in a tool spindle can cause vibrations in the tool spindle, the order of which (related to tool rotation) is determined by the bearing order. These bearing orders are generated by the bearing design and are often available from the bearing manufacturer. Therefore, vibrations measured during a test cycle can be directly attributed to worn bearings. These vibrations can be measured, for example, by an accelerometer on the workpiece spindle housing. The vibrations propagate from the machining process to the workpiece and manifest themselves as periodic deviations (ripples / waves) on the tooth surface. After installation in a gear train, these ripples manifest as noise excitations as the workpiece teeth roll on the mating gears. The order of these noise excitations related to the rotation of the workpiece in the gear train can be easily calculated based on the above considerations. In this way, it is possible to calculate how worn bearings will affect the noise spectrum of the gear train.

[0085] The calculated spectrum in Figure 5 only allows for no quantitative statement regarding the expected signal strength. However, by visualizing how this expected spectrum changes from one test cycle to another, an excellent estimate of the perturbation order and how it changes can be made. This is illustrated below with reference to Figure 6. Figure 6 shows a plot where the expected spectral strength I in an EOL test bench at a specific order (here, order 52) is plotted as a function of the number of workpieces processed by the machine. It can be seen that the expected noise strength increases strongly over time. This increase can be quantified by fitting a suitable regression function (here, a quadratic regression function), and depends on whether the determined regression parameters trigger a suitable action, such as issuing a warning signal. The temporal evolution of the expected signal strength at different EOL orders can also be visualized in a suitable manner. This allows even inexperienced users to interpret noise problems.

[0086] [Identifying the source of disturbance based on moving links]

[0087] The reverse is also possible: if the EOL spectrum has been determined by measurements of the EOL test bench, the above considerations can be used to estimate which machine axes and (if necessary) which components of the machine axes have caused disturbance orders in the measured EOL spectrum. For this purpose, the machine axis orders corresponding to the disturbance orders in the measured EOL spectrum are inverted and calculated, and components whose expected disturbance orders in the spectrum of the machine axes correspond to the orders calculated in this way are searched. This procedure can be easily automated.

[0088] [Procedures lacking knowledge of motion linkages]

[0089] If the motion linkage of the drive component is not known or cannot be used for calculation for other reasons, it may be possible to use machine learning (ML) algorithms to predict the signal strength or identify the source of the disturbance at certain EOL disturbance levels.

[0090] This scenario will be explained below with reference to Figure 7. This is a highly simplified schematic sketch of an Artificial Neural Network (ANN). In this example, the network has only three inputs and two outputs, and only one hidden layer. In practice, ANNs will typically have more inputs, outputs, and hidden layers.

[0091] In this example, state data is fed into the ANN at the input, each representing the vibration tendency of one of the machine's axes B, Y, and Z. The ANN then calculates the predicted EOL spectrum data in the form of the expected spectral intensities for two specific EOL orders (here, orders 52 and 59).

[0092] The ANN was previously trained using training data. Figure 8 shows a partial example of this type of training data. Figure 8 shows a graph in which, on one hand, state data determined from numerous test cycles of the machine is entered. On the other hand, EOL data in the form of spectral intensity at steps 52 and 59 is entered, obtained from EOL measurements on the gear train. In these gear trains, the workpiece is machined while it is in the state for which the state data was obtained (i.e., exactly before and / or after each test cycle). The graph contains many columns of this type. These columns can be obtained from the database DB in Figures 1 and 3. The ANN has been trained using this data in a manner known to the data itself. Therefore, it is possible to reliably predict which machine states (represented by the state data) will result in which EOL intensity at the mentioned steps.

[0093] The reverse is also possible: the input variables of the corresponding ANN can be EOL data, and the output data can be predicted state data.

[0094] Of course, the above example is highly simplified, but it illustrates the basic approach. Instead of ANNs, other types of machine learning algorithms can be used, especially other well-known classifiers.

[0095] [Results and Visualization Output]

[0096] The visualization of these analytical results can be performed independently of any user's computer via a web browser. Other evaluation measures can also be implemented in a corresponding platform-independent manner. This even facilitates analysis remotely. Specifically, the status of any machine can be examined in detail via the cloud from any mobile device.

[0097] In addition, it is conceivable that if there is a situation requiring intervention, the corresponding message will be automatically sent via SMS, push message or email.

[0098] 1: Grinding machine 2: EOL Test Stand 11: Machine tool 12: Tool rack 13: Axial sliding component 14: Grinding head 15: Tool Spindle 16: Grinding worm gear 18: Sensor 19: Sensor 20: Workpiece carrier 21: Workpiece spindle 22: Tail Seat 23: Workpiece 24: Engagement probe 30: Dressing device 33: Repair tools 40: Machine controller 41: Axis Module 42: Controlling the computer 43: Control Panel 44: Monitoring device 45: Server 46: Database Server 47: Web server 48: Terminal device 51: Accelerometer 52: Position sensor / step 53: Current sensor 54: Accelerometer 59: level 451: Processor 452: Storage Media A: Axis B: Tool axis C1: Workpiece axis C2: Workpiece axis C3: Axis of rotation DB: Database I: Expected Spectral Intensity X: Radial feed direction / Constant radial Y: Shift direction / offset axis Z: Axial feed direction

Claims

1. A method for monitoring the state of a gear cutting machine having a plurality of machine axes, comprising the steps of: a) performing a test cycle, wherein during the test cycle, at least a portion of the machine axes is systematically actuated and associated machine measurement data is obtained; b) performing a spectral analysis of the machine measurement data, wherein machine spectral data is calculated from the machine measurement data; and c) determining predicted EOL spectral data based on the machine spectral data, wherein the predicted EOL spectral data indicates the order in an EOL spectrum expected to be present when a workpiece processed by the gear cutting machine is mounted in a gear train and rolls on a mating gear in the gear train.

2. The method of request item 1, comprising: d) outputting the predicted EOL spectrum data or at least one quantity derived therefrom.

3. The method of claim 1 or 2, wherein determining the predicted EOL spectrum data includes applying a propagation factor to the machine spectrum data, the propagation factor depending on a motion link between the machine axis and the workpiece in determining the machine spectrum data.

4. The method of request item 1 or 2, wherein the predicted EOL spectrum data is determined individually for each actuated machine axis.

5. The method of claim 1 or 2, wherein steps a) to c) are repeated a number of times, wherein the workpiece is machined with the gear cutter between the test cycles and the test cycles are performed during a processing pause when the machining tool is not engaged with a workpiece, and wherein the development of one of the predicted EOL spectrum data is visualized and / or analyzed as a function of the test cycle performed, the workpieces processed, or the time, specifically visualized and / or analyzed by a regression analysis.

6. The method of claim 1 or 2, wherein reference machine spectrum data is available for a plurality of reference machines, the reference machine spectrum data having been determined by a plurality of reference test cycles performed on such reference machines, wherein a predicted reference EOL spectrum data is determined from the reference machine spectrum data, and wherein the predicted EOL spectrum data determined based on the machine spectrum data of the monitored gear cutter is compared with the predicted reference EOL spectrum data or a quantity derived therefrom.

7. The method of claim 6, wherein the method includes a statistical analysis of one of the predictive reference EOL spectral data.

8. A method for monitoring the state of a gear cutting machine having a plurality of machine axes, comprising the steps of: a) performing an EOL test on a gear train comprising a workpiece processed by the gear cutting machine, wherein in the EOL test, the workpiece in the gear train rolls on a mating gear, and determining associated EOL measurement data; b) performing a spectral analysis on the EOL measurement data, wherein calculating EOL spectral data from the EOL measurement data; and c) determining predicted state data based on the EOL spectral data, wherein the predicted state data for at least one machine axis indicates which orders of that machine axis are consistent with the calculated EOL spectral data.

9. The method of request item 8 includes: d) outputting the predicted state data or at least one quantity derived therefrom.

10. The method of claim 8 or 9, further comprising: e) performing a test cycle in which at least a portion of the machine axes is systematically actuated and associated machine measurement data is obtained; f) performing a spectral analysis of the machine measurement data, wherein machine spectral data is calculated from the machine measurement data; and g) determining predicted EOL spectral data based on the machine spectral data, wherein the predicted EOL spectral data indicates the order in an EOL spectrum expected to occur when a workpiece processed by the gear cutter is mounted in a gear train and rolls on one of the mating gears in the gear train, wherein determining the predicted state data includes comparing the EOL spectral data calculated from the EOL measurement data with the predicted EOL spectral data.

11. A method for establishing a training dataset for a machine learning algorithm, wherein the machine learning algorithm is an algorithm for monitoring the state of a gear cutting machine having a plurality of machine axes, wherein the machine learning algorithm comprises: a) performing a test cycle in which at least a portion of the machine axes is systematically actuated and associated state data is determined by measurement; b) machining at least one workpiece with the gear cutting machine when the gear cutting machine is in a state corresponding to the state data; c) mounting the machined workpiece in a gear train; d) performing an EOL test on the gear train, wherein in the EOL test, the workpiece in the gear train rolls on a mating gear and associated EOL data is determined; e) storing the state data and the corresponding EOL data in the training dataset; f) repeating steps a) to e) for a plurality of test cycles and machined workpieces, wherein the workpieces have the same nominal geometry and are machined in the same machining state.

12. A method for training a machine learning algorithm, wherein the machine learning algorithm is trained using the training dataset as described in claim 11.

13. A method for monitoring the state of a geared machine having a plurality of machine axes, characterized in that the method uses a machine learning algorithm trained by the training dataset as claimed in claim 11.

14. The method of claim 13, wherein the machine learning algorithm has state data of the gear cutting machine as input variables and predicted EOL data as output variables, the method comprising: a) performing a test cycle, wherein in the test cycle, at least a portion of the machine axes is systematically actuated and associated state data is determined by measurement; and b) determining the predicted EOL data based on the state data by feeding the state data to the trained ML algorithm as input variables.

15. The method of claim 13, comprising: c) outputting the predicted EOL spectral data or at least one quantity derived therefrom.

16. The method of claim 13, wherein the machine learning algorithm has EOL data as input variables and predicted state data of the gear cutting machine as output variables, the method comprising: a) performing an EOL test on the gear train, wherein in the EOL test, it is determined that the workpiece in the gear train rolls on the mating gears and associated EOL data is determined; and b) determining the predicted state data based on the EOL data by feeding the EOL data to the trained ML algorithm as input variables.

17. The method of request item 16, comprising: c) outputting the state data of the prediction or at least one quantity derived therefrom.

18. The method of any one of claims 13 to 17, wherein the machine learning algorithm is a classification algorithm, specifically an artificial neural network or a support vector machine or a random forest.

19. The method of any one of claims 13 to 17, wherein the state data is related to a state of a machine axis with respect to its vibration behavior and includes machine spectrum data calculated by spectral analysis of machine measurement data, and / or wherein the EOL data is related to the noise behavior of the gear train and includes EOL spectrum data calculated by spectral analysis of EOL measurement data.

20. An apparatus for monitoring the state of a gear cutting machine having a plurality of machine axes, comprising a processor (451) and a storage medium (452) storing a computer program that, when executed on the processor, causes the following steps to be performed: receiving state data determined by a test cycle of the gear cutting machine, wherein at least a portion of the machine axes has been systematically actuated during the test cycle, and the associated state data has been determined by measurement; and based on the state data, determining predicted EOL data related to noise behavior of a gear system comprising a workpiece processed by the gear cutting machine.

21. An apparatus for monitoring the state of one of a gear cutting machine having a plurality of machine axes, comprising a processor (451) and a storage medium (452) storing a computer program that, when executed on the processor, causes the following steps to be performed: receiving EOL data determined by an EOL test on one of a gear trains containing a workpiece processed by the gear cutting machine, wherein in the EOL test, the workpiece in the gear train rolls on a mating gear, and determining the associated EOL data; and based on the EOL data, determining predictive state data related to a state of at least one machine axis in terms of its vibration behavior.

Images