Method for monitoring the condition of a hard finishing machine, in particular a grinding machine

Abstract

The invention relates to a method for monitoring the condition of a hard finishing machine, in particular a grinding machine, wherein the hard finishing machine has a number of NC-controlled axes that are actuated during the machining of a workpiece, wherein at least one reference run is performed on at least one axis to assess the condition of the hard finishing machine, during which the response of the axis to a drive signal is measured and evaluated. To enable improved condition monitoring of the hard finishing machine, the invention provides that the reference run includes a phase in which the axis is accelerated and / or decelerated.

Description

[0001] The invention relates to a method for monitoring the condition of a hard finishing machine, in particular a grinding machine, wherein the hard finishing machine has a number of NC-controlled axes that are actuated during the machining of a workpiece, wherein at least one reference run is carried out on at least one axis for the assessment of the condition of the hard finishing machine, in which the reaction of the axis to a drive signal is measured and evaluated.

[0002] A hard finishing machine, especially a gear or profile grinding machine, is subject to wear and tear due to operation, so machine maintenance must be carried out at appropriate times and machine components replaced if necessary. The same applies, for example, in the event of an unexpected event (e.g., crash, overload). It is essential to prevent workpiece production from taking place while the machine is already out of proper working order due to wear or damage.

[0003] Therefore, in practice, it is common to periodically perform reference runs of the machine to check its status and, in particular, to assess wear. During these runs, the machine axes (linear and rotary) are moved outside of normal operating conditions, and the response to predefined control signals is recorded. This allows verification of whether the machine components involved are still in perfect working order or whether changes in the response to predefined control signals have already been detected due to wear or an unexpected event (e.g., a crash). In the latter case, appropriate machine maintenance is then required.

[0004] The cyclical reference runs provide a very accurate picture of the machine's condition. Precise results are obtained because predefined movements can be performed under similar conditions during the reference runs, thus minimizing the impact of disruptive factors on the system.

[0005] A disadvantage, however, is that the frequency with which changes to the machine can be detected depends on the chosen period in which these reference runs are performed. In practice, efforts are made to perform reference runs infrequently due to the time required (since no workpiece machining can take place during this time), which then leads to relatively long periods during which the machine is not checked (long intervals between reference runs). Furthermore, longer reference runs can potentially lead to problems with the machine's temperature coefficient.

[0006] The invention is based on the objective of further developing a method of the type mentioned above in such a way as to enable improved condition monitoring of the hard finishing machine and, in particular, its machine components, especially the linear and rotary axes. This should allow for close-meshed machine monitoring with as little loss of machine capacity as possible. Furthermore, the aim is to detect potential causes of wear more precisely.

[0007] The solution to this problem by the invention is characterized in that the reference run includes a phase in which the axis is accelerated and / or decelerated.

[0008] Preferably, the response of the axis to the drive signal is measured and the measured response is compared with expected signal curves, whereby a statement is derived from the comparison as to whether the axis is in a proper condition.

[0009] The aforementioned predetermined reactions are preferably those that were present when the axle was in proper working order (or were determined when the axle was in proper working order). Combinations of several reactions resulting from a proper working order can also be considered. Such reactions may also originate from and have been recorded on other machines.

[0010] The aforementioned comparison is preferably performed using a machine learning or deep learning concept. The machine learning or deep learning is preferably performed using at least one autoencoder.

[0011] The detected and evaluated response of the axis to the drive signal can, for example, and preferably, be a control deviation of the NC axis that occurs during the control of the machine axis, or it can be an acceleration / vibration detected by an accelerometer, and / or it can be a current with which a machine axis is driven. Other signals that provide significant information about the acceleration or deceleration process can also be considered, and such signals can be taken into account alternatively or additively.

[0012] The reference run is preferably performed when there is no engagement of the tool of the hard finishing machine with the workpiece.

[0013] The reference run can include at least one acceleration phase and / or deceleration phase, as well as a phase of constant speed of the axis.

[0014] The reference run can also be performed as a test run, which takes place during a machining phase of the workpiece, beginning with the workpiece being placed in the hard finishing machine and ending with the removal of the machined workpiece from the hard finishing machine. In this case, it can be stipulated that the test run takes place when the tool of the hard finishing machine is not engaged with the workpiece.

[0015] The test run preferably takes place during an idle stroke, in which the tool is moved relative to the workpiece without the tool engaging with the workpiece.

[0016] With regard to the aforementioned comparison of the measured response of the axis to the drive signal with previously determined responses, a preferred embodiment of the invention provides for an assessment of whether a deviation from a permissible tolerance band is to be expected. In this respect, a trend analysis is performed, incorporating previously recorded values ​​to determine whether the machine axis is about to leave a permissible operating range. Artificial intelligence methods are also very well suited for this purpose, enabling this to be done automatically.

[0017] The method is preferably a gear grinding process.

[0018] Thus, an efficient assessment of the machine's wear condition is achieved by evaluating signal profiles recorded when an axis undergoes at least one acceleration or deceleration phase. These acceleration and deceleration phases are performed during a reference run or through idle strokes during machine operation. Established artificial intelligence methods, particularly machine learning and deep learning, are preferably employed to evaluate the recorded signals.

[0019] The proposed procedure allows for improved detection of machine malfunctions or the current wear condition of the machine.

[0020] In addition to the known evaluation of constant speed operation from a reference run, the invention integrates at least one acceleration and / or one deceleration phase (from a reference run and / or from idle strokes during the process) into the evaluation. These acceleration and deceleration phases represent nonlinearities significantly better than constant speed operation, such as those that occur with play or backlash. This allows for faster and earlier detection of incipient wear. In particular, acceleration from standstill or deceleration to standstill in linear axes serves to accurately detect such defects (nonlinearities).

[0021] Furthermore, the more comprehensive evaluation of the reference run offers the advantage of a holistic assessment of the entire system, consisting of individual mechanical components of the axis, controller settings, drives, etc. Among other things, current spikes, stiffness, and lubrication conditions can be detected more effectively during the acceleration phases.

[0022] This is particularly important for components with gears for electromobility due to increased quality requirements or tight workpiece tolerances.

[0023] Statistical or parameter-based approaches are not optimal for evaluating nonlinear phases. Due to the complexity of the problem, the entire signal waveforms and / or combinations of signal waveforms are advantageous for evaluation. Machine learning and deep learning methods are particularly suitable for this purpose. These methods can also capture complex changes in the system.

[0024] Regarding the nomenclature used above, it should be noted that a "test run" refers to the described monitoring of a machine axis during the regular machining process of a workpiece - although this only takes place for partial machining processes in which the tool does not engage with the workpiece - while a "reference run" is a separate test run of the (complete) machine axes, which is carried out outside the machining process of a workpiece.

[0025] A test run can preferably take place during a dressing phase in which the tool is profiled, and / or during a phase in which the workpiece is aligned for machining in the hard finishing machine, and / or during a phase in which a balancing operation is carried out, and / or in a phase in which a loading or unloading operation is carried out.

[0026] According to the proposed concept, the condition of the machine, and especially its components, can be recorded during a workpiece machining cycle, thus enabling close and continuous monitoring. The time-consuming and therefore only spot-checked reference runs no longer need to be performed cyclically, but can be carried out as needed.

[0027] Including at least one acceleration phase or at least one deceleration phase in the evaluation has shown that a much better assessment of the machine's status is possible than with the previously known evaluation when the machine axis is in constant motion.

[0028] The monitoring quality increases due to the more frequent inspections, and downtime can be advantageously reduced. This results in a reduction in the number of improperly machined workpieces (due to worn or faulty machine components). Furthermore, root cause analysis is significantly simplified, providing a solid foundation for preventative machine maintenance (the "predictive maintenance" concept).

[0029] Another advantage is that causes, especially for sudden wear (for example, the ingress of dirt or other abrasive media into bearings or ball screws, or a crash or overload of the machine), can be identified much more easily and assigned much more accurately.

[0030] Furthermore, it is advantageous that the availability of the machine for the effective processing of the workpiece can be increased, since otherwise available downtime can be used for test runs and thus unplanned machine downtime can be reduced or avoided with the described procedure.

[0031] Finally, it is advantageous that critical temperature fluctuations of the machine can be avoided.

[0032] The drawing shows an embodiment of the invention. Fig. Figure 1 shows the course of a control error over time during the acceleration of a machine axis of a gear grinding machine, showing the new state (curve a), the beginning of wear (curve b) and the worn state (curve c). Fig. Figure 2 schematically shows the structure of an autoencoder, which is used for the evaluation of a captured signal within the framework of the “machine learning” concept. Fig. Figure 3 shows, as an example, (top) the original data input into the autoencoder, (middle) the encoded signal and (bottom) the decoded signal determined by the autoencoder, and Fig. Figure 4 schematically shows the composition of individual operating phases of the machine into models that can be analyzed.

[0033] In Fig. Figure 1 shows the control deviation of a machine axis of a gear grinding machine over time. The control deviation is the discrepancy between the specified movement and the actual movement of the machine axis. With progressive wear (and thus with increasing backlash and stiffness), it becomes increasingly difficult for the machine control system to bring the actual position as close as possible to the target position. Consequently, the control deviation can be used as a measure of the wear condition of the machine or of the components that influence the movement of the relevant machine axis.

[0034] The graph shows the resulting control error (in mm) for a period of 0.3 s in which the machine axis is accelerated from rest with a given acceleration.

[0035] Curve a represents the curve achieved when the machine is new. As wear and tear gradually begins, curve b results. Finally, when the machine is worn out, curve c is observed.

[0036] The key point is that particularly meaningful information about the wear condition of the axle can be obtained if, instead of simply maintaining a constant speed (constant speed of the machine axle) as before, the machine axle is given a predetermined acceleration to maintain, as shown.

[0037] As can be seen from the in Fig. As can be seen from the curve of the control deviation over time shown in Figure 1, a fairly precise statement can be made about the wear condition of the machine based on the curve's shape.

[0038] The evaluation of the recorded curve of the course of the control deviation over time, i.e., the comparison with previously recorded courses of the curve when the machine was new, solely by statistical methods or by means of characteristic value calculations, sometimes does not provide a sufficiently good result.

[0039] To account for the changes during the acceleration phase (as in Fig. 1 shown) and / or to evaluate in a deceleration phase, “machine learning” methods or “deep learning” methods are advantageously used.

[0040] A preferred application here is the use of an autoencoder. This consists of several specially arranged layers of a neural network designed to learn efficient codings, which can then be used to detect anomalies. The neural network is trained, for example, with data from reference runs that have been clearly identified as "correct." If data from a reference run with an unknown state (during the lifetime of the machine or axle) is then fed into the autoencoder's input, it attempts to reconstruct the signal based on its previously trained state or the learned coding. Error values ​​(such as "Mean Absolute Error" and "Mean Squared Error") are calculated between the input and output signals.When certain thresholds for these calculated parameters are exceeded, a warning about the axle's condition is issued, indicating that a critical wear condition has been reached. Furthermore, this approach allows warnings to be issued if a trend in the machine's condition is detected. For example, if several tests within a specific time interval show an increase in defined measured values, this can indicate impending wear, which the system can then signal.

[0041] The reference signals can and must be considered separately for different machine series. All algorithms are based on a sufficiently large and informative dataset. These datasets are derived from numerous measurement series conducted on machines in the field, in-house, and on specially designed test benches.

[0042] Therefore, evaluations using artificial intelligence are preferred, as described in the Fig. 2 and Fig. 3 is indicated by way of example.

[0043] In Fig. Figure 2 shows the basic structure of an autoencoder. It schematically illustrates how an autoencoder works to obtain a reconstructed input from input values. This reconstructed input allows conclusions to be drawn about how much the input values ​​have changed compared to predefined values, thus enabling the detection of relevant deviations.

[0044] The number of neurons (these are the circular symbols in Fig. 2) The number of neurons is cyclically reduced via an "Encoder Hidden Layer 1" and an "Encoder Hidden Layer 2" until it reaches the "Code Layer." The encoded signal can then be accessed there. From the "Code Layer" to the output layer ("Reconstructed Input"), the number of neurons is cyclically increased again from "Decoder Hidden Layer 1" via "Decoder Hidden Layer 2."

[0045] In Fig. Figure 3 illustrates this using a specific application example. The figure shows example signals at the different layers. The input signal (top representation in Fig. 3: Original data input) is continuously reduced in the number of samples by the "Encoder Hidden Layer" of the neural network until the encoded signal ("Code Layer") is reached (see middle illustration in Fig. 3: Encoded Signal). Subsequently, the "Decoded Hidden Layer" reduces the signal back to the original number of samples ("upsampled"; see lower illustration in Fig. 3: Decoded Signal).

[0046] Ideally, the autoencoder can reconstruct the signal flawlessly. However, if signals are fed into the autoencoder that deviate from the shape of the trained signals, the autoencoder cannot reconstruct the signal flawlessly. The calculated error values ​​increase as the deviation from the input signal (due to wear and tear) grows.

[0047] The system can then conclude that wear has occurred in the machine or the machine axis and issue a corresponding notification to the machine operator.

[0048] In Fig.Figure 4 schematically illustrates how models can be created, which then form the basis for the evaluation described above. It shows several measuring strokes (H1, H2, H3, H4) for one axis of the machine, each with an acceleration phase BP, a constant speed phase KP, and a deceleration phase AP. Different combinations of the machining components can be used for model creation.

[0049] According to a “Model a”, the acceleration phase BP, the constant phase KP and the deceleration phase AP are considered and, for the purpose of model building, the respective phases of the four strokes are either combined or each phase is used individually to form a model.

[0050] According to a “Model b”, the three phases BP, KP and AP are combined into one model.

[0051] According to a “Model c”, the three phases BP, KP and AP are combined, but considered separately for each hub.

[0052] According to a “Model d”, the data for the “Model c” are summarized for all strokes (one model for the entire reference run).

[0053] Of course, other models or types of neural networks can also be used to perform the described monitoring.

[0054] The described procedure allows essential information about the machine's condition to be determined, enabling deviations to be identified very quickly and reliably and appropriate measures to be taken if necessary; this prevents the production of workpieces that are scrap.

[0055] Furthermore, a comprehensive concept for predictive maintenance of machinery can be implemented.

[0056] This is particularly advantageous in cases of sudden wear, allowing for the timely detection of situations where, for example, machine components need to be replaced. If necessary, the optimal time for such a measure can be predicted by extrapolating the relevant parameters.

[0057] During phases of workpiece machining where the tool does not engage with the workpiece, in particular during tool dressing, balancing, swiveling of a rotary table or other machine movements necessary for the machining cycle, data from the individual axes are recorded.

[0058] These idle strokes (i.e., machine movements where the tool does not engage the workpiece) can include, in particular, an idle stroke during finishing (Y-axis) or roughing, the feed during dressing and grinding (X-axis), an idle stroke during grinding (Z-axis), or approaching the home position for balancing or centrifugal casting. Also well-suited are those phases of the machining process involving balancing, centrifugal casting, workpiece alignment, swiveling of a turret (rotary table) for dressing, and workpiece changes.

[0059] The recorded measurement signals may include, in particular, signals from an accelerometer mounted on a component of the machine, internal control signals (such as currents or control deviations), and other signals, if applicable, from existing sensors (for example, a current clamp and an encoder).

[0060] The processing of the signals preferably takes place on an industrial PC.

[0061] Signal progressions, evaluations, and analyses can be stored locally on the PC in a database. Alternatively, the data can be stored and analyzed in the cloud.

[0062] The proposed concept can be applied to specific assemblies, machines, or components, or across multiple machines. An approach spanning different machine types is also conceivable. A sufficiently large and adequate database is always necessary for any adaptation of the concept.

Claims

[1] Method for monitoring the condition of a hard finishing machine, in particular a grinding machine, wherein the hard finishing machine has a number of NC-controlled axes which are actuated during the machining of a workpiece, wherein for the assessment of the condition of the hard finishing machine at least one reference run is carried out on at least one axis in which the response of the axis to a drive signal is measured and evaluated, characterized by , that the reference run includes a phase in which the axle is accelerated and / or decelerated. [2] Method according to claim 1, characterized by , that the response of the axis to the drive signal is measured and the measured response is compared with expected signal curves, whereby a statement is derived from the comparison as to whether the axis is in a proper condition. [3] Method according to claim 2, characterized bythat the previously determined reactions are those that occurred when the axis was in proper working order. [4] Method according to claim 2 or 3, characterized by that the comparison is carried out using a machine learning concept or a deep learning concept. [5] Method according to claim 4, characterized by that the machine learning concept or the deep learning concept is used with the aid of at least one autoencoder. [6] Method according to any one of claims 1 to 5, characterized by , that the detected and evaluated response of the axis to the drive signal is a control deviation of the NC axis that occurs during the control of the machine axis, or is an acceleration detected by an accelerometer, and / or is a current with which a machine axis is driven. [7] Method according to any one of claims 1 to 6, characterized by, that the reference run takes place when there is no engagement of the tool of the hard finishing machine with the workpiece. [8] Method according to any one of claims 1 to 7, characterized by that the reference run includes at least one acceleration phase and / or deceleration phase, as well as a phase of constant speed of the axis. [9] Method according to any one of claims 1 to 8, characterized by , that the reference run is carried out as a test run, which takes place during the period of a machining phase of the workpiece, which begins with the insertion of the workpiece into the hard finishing machine and ends with the removal of the machined workpiece from the hard finishing machine. [10] Method according to claim 9, characterized by , that the test run takes place when there is no engagement of the tool of the hard finishing machine with the workpiece. [11] Method according to claim 10, characterized by, that the test run takes place during an idle stroke, in which the tool is moved relative to the workpiece without the tool engaging with the workpiece. [12] Method according to any one of claims 2 to 11, characterized by , that when comparing the measured response of the axis to the drive signal with previously determined responses, it is assessed whether a departure from a permissible tolerance band is to be expected. [13] Method according to any one of claims 1 to 12, characterized by that it is a gear grinding process.

Images