Method for monitoring state of hard finishing machine, in particular grinding machine

Abstract

The invention relates to a method for monitoring the state of a hard finishing machine, in particular a grinding machine, the hard finishing machine having a plurality of numerical control shafts which are actuated during the machining of a workpiece, the numerical control shafts being actuated during the machining of the workpiece for the purpose of evaluating the state of the hard finishing machine. At least one reference operation is carried out on at least one shaft, and the response of the shaft to the drive signal is measured and evaluated during the reference operation. In order to be able to improve the monitoring of the state of the hard finishing machine, the reference operation according to the invention comprises a phase in which the shaft is accelerated and / or braked.

Description

Technical Field

[0001] The present invention relates to a method for monitoring the condition of a hard finishing machine, particularly a grinding machine, wherein the hard finishing machine has a plurality of CNC axes that are actuated during workpiece machining, wherein, for the purpose of evaluating the condition of the hard finishing machine, at least one reference run is performed on at least one axis, during which the response of the axis to a drive signal is measured and evaluated. Background Technology

[0002] EP 4130902 A1 discloses a method for starting a production machine with a position-controlled axis, wherein the control device stores the maximum stroke range, the maximum stroke speed of the axis, and the maximum drive power as boundary conditions for identification operation. Identification operation is performed using these parameters, and the relevant parameters are monitored during the identification operation. Identification is based on a model with the input parameters.

[0003] EP 3176657 A1 discloses a method for determining the stiffness of a transmission system when a machine part moves linearly along a linear guide of a machine tool. A constant acceleration is assigned to the machine part, and the difference between a specified position and a measured position of the machine part is determined.

[0004] WO 2023 / 061851 A1 discloses a method for monitoring the condition of a machine tool, which can be used to predict the generation of interference noise in a gearbox containing a workpiece being machined. In this method, machine spectrum data is calculated based on machine measurement data through spectrum analysis to determine at which frequencies or orders the machine components are periodically excited while the machine axes are running.

[0005] DE 102023101527 A1 discloses a method for monitoring the state of an electric drive system, wherein process data and drive data are recorded using identification features. This is intended to enable better detection of the target operating state.

[0006] Hard finishing machines, especially gear grinders or profile grinders, are prone to wear and tear during operation. Therefore, machine maintenance must be performed at appropriate intervals, and machine parts must be replaced when necessary. This also applies in the event of unexpected events (such as machine failure or overload). Importantly, production should be avoided when the machine is unable to operate normally due to wear or damage.

[0007] Therefore, it is common practice to periodically perform baseline runs on the machine to check its condition and, in particular, assess wear. During these runs, the machine axes (linear / rotary axes) are moved outside the machine's actual operating range, and specified control response signals are recorded. This allows for checking whether the machine components involved are still in perfect condition, or whether changes in the response to specified control signals due to wear or unexpected events (e.g., a system crash) have been detected. In the latter case, appropriate machine maintenance is required.

[0008] Cyclic benchmark runs reflect a very accurate machine condition. Because benchmark runs can be performed under similar conditions with predetermined motion, accurate results are obtained, meaning that disturbances have minimal impact on the system.

[0009] However, a drawback here is that the frequency at which machine changes can be detected depends on the cycle selected for the baseline run. In practice, due to the time required, baseline runs are avoided as much as possible (because no workpieces can be processed during these runs), which would result in the machine being unchecked for a considerable period (long intervals between baseline runs). Long baseline run times can also cause problems with the machine's temperature profile. Summary of the Invention

[0010] The present invention aims to further develop methods of the type described above, thereby improving the condition monitoring of hard-finishing machines, particularly their machine components, especially linear and rotary axes. This should enable close-up monitoring of the machine with minimal loss of machine capability. Another objective is to more accurately detect potential causes of wear.

[0011] The present invention provides a solution to this problem, characterized in that the reference operation includes a phase in which the shaft is accelerated and / or braked, wherein, in order to assess the wear condition of the hard-finishing machine, the response of the shaft to the drive signal is measured, and the measured response is compared with the expected signal curve that would appear when the shaft is in the proper condition, wherein a statement about whether the shaft is in the proper condition is made from the comparison.

[0012] Therefore, the predetermined reaction mentioned is the reaction that occurs when the shaft is in its normal state (or the reaction determined when the shaft is in its normal state). Combinations of several reactions derived from the normal state can also be considered. Such reactions may also originate from other machines and have been recorded on those machines.

[0013] Preferably, the comparison is performed using machine learning or deep learning concepts. Preferably, machine learning or deep learning is performed using at least one autoencoder.

[0014] The measured and evaluated axis response to the drive signal can be (e.g., and preferably) a control error occurring during machine axis control, or it can be acceleration / oscillation recorded by an accelerometer, and / or it can be the current used to drive the machine axis. Other signals that provide important information about the acceleration or deceleration process may also be considered, whereby these signals may be optionally or additionally taken into account.

[0015] Preferably, the benchmark run is performed when there is no engagement between the cutting tool and the workpiece in the hard finishing machine.

[0016] In addition to at least one acceleration phase and / or deceleration phase, the benchmark run may also include a constant speed phase for the shaft.

[0017] This benchmark run can also be performed as a trial run during the machining phase of the workpiece. The benchmark run begins when the workpiece is inserted into the hard finish machine and ends when the machined workpiece is removed from the hard finish machine. In this case, it can be provided that the trial run is performed when there is no engagement between the cutting tool and the workpiece on the hard finish machine.

[0018] Preferably, the test run is performed during the idle stroke of the tool moving relative to the workpiece but not engaging with the workpiece.

[0019] Regarding the comparison above between the measured shaft response to the drive signal and the predetermined response, a preferred embodiment of the invention provides an assessment of whether the machine axis is expected to exceed the permissible tolerance zone. In this regard, trend analysis is performed, taking into account previously recorded values, to determine whether the machine axis is about to leave the permissible operating range. Artificial intelligence methods are also well-suited for automating this operation.

[0020] Preferably, the method is used on a gear grinding machine.

[0021] Thus, by evaluating the signal curves determined when the shaft passes through at least one acceleration phase or at least one deceleration phase, the wear condition of the machine can be effectively assessed. These acceleration or deceleration phases are performed during baseline operation or via a no-stroke operation during machine operation. To evaluate the recorded signals, known artificial intelligence methods, particularly machine learning or deep learning, are preferably used.

[0022] The proposed process can improve the detection of machine failures and the current wear and tear of the machine.

[0023] In addition to known evaluations of constant motion based on a reference run, this invention integrates at least one acceleration or deceleration phase (from the reference run and / or from the idle travel during the process) into the evaluation. These acceleration and deceleration phases are more representative of nonlinear characteristics than constant motion, such as motion when components become loose or gaps appear. This allows for faster detection of initial wear at earlier stages. In particular, the acceleration or deceleration of a linear axis from rest to rest helps to correctly detect such errors.

[0024] Another advantage is that a more comprehensive evaluation of baseline operation allows for a holistic assessment of the entire system, comprising the individual mechanical components of the shaft, controller settings, drives, and more. Furthermore, it enables more effective detection of current peaks, hysteresis, and lubrication conditions during the acceleration phase.

[0025] This is especially important for geared components in electric vehicles, as these components have higher quality requirements and stricter workpiece tolerances.

[0026] For evaluating nonlinear phases, statistical or feature-based methods are not ideal. Due to the complexity of the problem, complete signal curves and / or combinations of signal curves are advantageous for evaluation. Machine learning or deep learning methods are better suited for evaluation. These methods can also detect complex variations in the system.

[0027] Regarding the terminology used above, it should be noted that "trial run" refers to the described monitoring of the machine axis during the routine machining process of the workpiece, although this applies only to the machining subprocess where the tool is not engaged with the workpiece, while "baseline run" is a separate trial run of the (complete) machine axis outside of the workpiece machining process.

[0028] Trial runs may be particularly preferred in the following stages: the tool dressing stage, and / or the workpiece alignment stage in a hard finishing machine, and / or the balancing stage, and / or the loading or unloading stage.

[0029] Based on the proposed concept, it is possible to record the state of machines, and especially machine parts, during workpiece machining cycles, thereby enabling intensive monitoring unaffected by time. Therefore, time-consuming baseline runs that were previously performed only occasionally no longer need to be executed cyclically, but can be performed on demand.

[0030] Incorporating at least one acceleration phase or at least one deceleration phase into the assessment demonstrates that the machine condition can be better evaluated compared to previously known assessments based on constant motion of the machine axis.

[0031] Timely, meticulous planning improves monitoring quality and minimizes non-productive time. This reduces the number of improperly processed parts (due to wear and tear and malfunctions of machine components). Furthermore, failure root cause analysis becomes easier, providing a solid foundation for preventative machine maintenance and repair (the concept of "predictive maintenance").

[0032] Another advantage is that it makes it easier to identify and more accurately pinpoint the cause of sudden wear (such as dirt or other abrasive media entering the bearing or ball screw drive, or machine malfunction or overload).

[0033] Another advantage is that by using the downtime that would otherwise be available for trial runs, the availability of the machine to effectively process workpieces can be improved, thereby reducing or avoiding unplanned machine downtime through the process.

[0034] Finally, another advantage is that it can prevent critical temperature fluctuations in the machine. Attached Figure Description

[0035] Figure 1 The diagram illustrates the change in control difference over time during the acceleration of the machine shaft in a gear grinding machine, indicating the new machine state (curve a), wear initiation (curve b), and wear state (curve c).

[0036] Figure 2 The diagram schematically illustrates the structure of an autoencoder used to evaluate detected signals in the concept of "machine learning."

[0037] Figure 3 The original data input to the autoencoder (top), the encoded signal (middle), and the decoded signal determined by the autoencoder (bottom) are illustrated by example.

[0038] Figure 4 This schematically illustrates a model that can be analyzed by comprising the various operational stages of a machine. Detailed Implementation

[0039] Figure 1 This diagram illustrates the time-varying control error of a machine axis on a gear grinder. Control error refers to the deviation between the intended movement and the actual movement of a machine axis. As wear intensifies (and consequently, the gradually increasing backlash and hysteresis), it becomes increasingly difficult to control the machine to make the actual position as close as possible to the desired position. Therefore, control error can be used as an indicator of the wear condition of the machine or those components affecting the movement of the relevant machine axis.

[0040] The figure shows the control deviation (in mm) produced when the machine axis accelerates from rest for 0.3 seconds with a specified acceleration.

[0041] Curve a represents the progress achieved when the machine is brand new. As the machine gradually begins to wear out, curve b appears. Finally, as the machine wears out, curve c appears.

[0042] Importantly, if the machine shaft is subjected to the specified acceleration shown and then must be operated in this manner (instead of operating at a constant speed as before), then particularly important information about the condition of the shaft wear can be obtained.

[0043] from Figure 1 The curve showing the control difference over time demonstrates that the curve pattern allows for a fairly accurate description of the machine's wear condition.

[0044] Simply using statistical methods or relying on eigenvalues ​​to evaluate the curve of the captured control difference over time (i.e., comparing it with the curve previously recorded when the machine was brand new) sometimes fails to provide sufficiently good results.

[0045] Advantageously, the acceleration phase can be evaluated using "machine learning" or "deep learning" methods (such as...). Figure 1 (as shown) and / or changes during the deceleration phase.

[0046] The preferred method here is to use an autoencoder. This autoencoder consists of a specially arranged neural network with multiple layers designed to learn efficient codes that can be used to detect anomalies. For example, the neural network is trained using data from a baseline run that has been explicitly classified as "normal." If data from a baseline run (during the machine's or shaft's lifespan) with an unknown state is subsequently fed into the autoencoder, it attempts to reconstruct the signal based on the previously trained state or the learned codes. Error values ​​(such as "mean absolute error" or "mean squared error") between the input and output signals are calculated. If certain thresholds for these calculated parameters are exceeded, a message is issued indicating the shaft's condition and that it has reached a critical wear state. This process also allows for message issuance upon detecting trends in machine condition. For example, if multiple tests over a certain time interval show an increase in a defined measurement, this can indicate initial wear, thus signaling it through the system.

[0047] For different machine families, reference signals can or must be considered separately. All algorithms are based on sufficiently large and significant datasets. These datasets are based on a large number of measurement sequences performed on machines in the field, internally, or on specially constructed test benches.

[0048] This means that AI-based evaluations are preferred, such as... Figure 2 and Figure 3 As shown.

[0049] Figure 2The basic structure of an autoencoder is shown. It provides a schematic diagram of how an autoencoder obtains a "reconstructed" input based on an "input" value, thus allowing conclusions to be drawn about the degree of change of the input value relative to a specified value, in order to identify any relevant deviations.

[0050] Neurons are reduced through a loop of "encoder hidden layer 1" and "encoder hidden layer 2". Figure 2 The number of neurons (circular symbols) is increased until the "code layer" is reached. Here, the encoded signal can be obtained. From the "code layer" to the output layer ("reconstructing the input"), the number of neurons increases again through "decoder hidden layer 1" via "decoder hidden layer 2".

[0051] Figure 3 This is illustrated with a concrete example. The diagram shows sample signals from different "layers." The input signal (passes through the "encoder hidden layer" of the neural network) Figure 3 The diagram above illustrates the process of continuously reducing the number of samples (raw data input) until the encoded signal ("code layer") is reached (see diagram above). Figure 3 The middle diagram shows the encoded signal. Then, through the "decoding hidden layer," the signal is converted back to the original number of samples ("upsampling"). Figure 3 The image below shows the decoded signal.

[0052] Ideally, an automatic encoder can reconstruct a signal without error. However, if the signal fed into the automatic encoder differs in form from the signal being learned, the encoder cannot reconstruct the signal without error. As the deviation from the input signal increases (due to wear), the calculated error value becomes larger.

[0053] The system can then detect wear in the machine or machine shafts and send corresponding messages to the machine operator.

[0054] Figure 4 The diagram schematically illustrates how to create a model, which is then used as the basis for the aforementioned evaluation. Multiple measurement strokes (H1, H2, H3, H4) of one axis of the machine are shown, each stroke having an acceleration phase BP, a constant speed phase KP, and a deceleration phase AP. The model can be formed using different combinations of machined parts.

[0055] According to "Model a", the acceleration phase BP, the constant speed phase KP, and the deceleration phase AP are considered. For modeling purposes, the corresponding phases of the four strokes are combined or each phase is used separately in a model.

[0056] Based on “Model b”, the three stages BP, KP and AP are combined into a single model.

[0057] Based on "Model c", the three stages BP, KP and AP are combined, but each trip is considered separately.

[0058] Based on "Model d", the data from "Model c" is summarized for all runs (one model is used for the entire baseline run).

[0059] Of course, other models or types of neural networks can also be used to perform the above monitoring.

[0060] The above process allows for the identification of critical information regarding the machine's condition, enabling the rapid and reliable detection of deviations and the implementation of appropriate corrective measures when necessary. This prevents the production of defective parts.

[0061] In addition, a comprehensive concept of preventative machine maintenance (“predictive maintenance”) can be implemented.

[0062] This is particularly advantageous in cases of sudden wear and tear, as it allows for the rapid identification of machine parts that need replacement, for example. If necessary, parameters can be deduced to predict the optimal timing for taking action.

[0063] During the machining of a workpiece where the tool is not engaged with the workpiece, especially during tool dressing, balancing, turntable rotation, or other machine movements required for the machining cycle, data from each axis is recorded.

[0064] These idle strokes (i.e., machine movements where the tool is not engaged with the workpiece) can specifically include: idle strokes during fine dressing (Y-axis) or rough dressing processes, feeds during dressing and grinding (X-axis) processes, idle strokes during grinding (Z-axis) processes, or movement to the starting position for balancing or rotation. They are also well-suited for various stages of the machining process, such as balancing, spinning, workpiece alignment, dressing on a rotating turret (turntable), and workpiece changes.

[0065] In particular, the recorded measurement signals can be signals from acceleration sensors arranged on machine parts, control internal signals (such as current or control differential), and other signals from existing sensors (e.g., from current measuring clamps and transmitters) (if necessary).

[0066] Preferably, the signal is processed on an industrial computer.

[0067] Signal curves, evaluations, and analyses can be stored in a local database on the computer. Data can also be stored and evaluated in the cloud.

[0068] The proposed concepts can be applied to specific components, machines, or parts, or even across multiple machines. Approaches covering different machine types are also possible. All adjustments to the concepts always require a sufficient and large database.

Claims

1. A method for monitoring the condition of hard finishing machines, particularly grinding machines, wherein, The hard finish machining machine has multiple CNC axes that are actuated during workpiece machining. For the purpose of evaluating the state of the hard finish machining machine, at least one reference run is performed on at least one axis, during which the response of the axis to drive signals is measured and evaluated. Its features are, The benchmark operation includes phases in which the shaft is accelerated and / or braked, wherein, in order to assess the wear condition of the hard finishing machine, the response of the shaft to the drive signal is measured, and the measured response is compared with an expected signal curve that would appear when the shaft is in an appropriate state, wherein a statement about whether the shaft is in an appropriate state is made from the comparison.

2. The method according to claim 1, characterized in that, The comparison is performed using machine learning or deep learning concepts.

3. The method according to claim 2, characterized in that, The machine learning concept or the deep learning concept is used in conjunction with at least one autoencoder.

4. The method according to claim 1, characterized in that, The measured and evaluated response of the axis to the drive signal is either a control difference that occurs during machine axis control of the CNC axis, or an acceleration detected by an accelerometer, and / or the current used to drive the machine axis.

5. The method according to claim 1, characterized in that, The reference run is performed when there is no engagement between the cutting tool and the workpiece in the hard finishing machine.

6. The method according to claim 1, characterized in that, In addition to at least one acceleration phase and / or deceleration phase, the baseline operation also includes a constant speed phase for the shaft.

7. The method according to claim 1, characterized in that, The benchmark run is performed as a trial run, which is conducted during the machining phase of the workpiece. The benchmark run begins when the workpiece enters the hard finishing machine and ends when the machined workpiece is removed from the hard finishing machine.

8. The method according to claim 7, characterized in that, The trial run is performed when there is no engagement between the cutting tool and the workpiece in the hard finishing machine.

9. The method according to claim 8, characterized in that, The test run is performed during the idle stroke when the cutting tool moves relative to the workpiece but is not engaged with the workpiece.

10. The method according to claim 1, characterized in that, When the measured response of the shaft to the drive signal is compared with a predetermined response, an assessment is made as to whether it is expected to exceed the permissible tolerance zone.

11. The method according to claim 1, characterized in that, The method is used on a gear grinding machine.

Images