Computer-implemented method, computer program product and apparatus for predicting wear of a machining tool

Abstract

The invention relates to a method for deriving an optimal point in time for replacing a tool, for which the wear of the tool must be derived. The solution is based on a time-dependent modeling of the wear between two wear measurement points, which are measured at the beginning and at the end of the life cycle of the machining tool, respectively. The point in time of the measurement is decisive, since the cutting process should not be impaired. The measurement point can be recorded by means of an arbitrary measurement system. Then, on the basis of the measurement values, a neural network is suitably trained, which can make a prediction about the respective degree of wear.

Description

Background Technology

[0001] In industrial manufacturing, a crucial component is the central controller, which is primarily composed of software. The purpose of this software is to control the tools in order to trigger the transport or processing of the workpieces to be manufactured.

[0002] A typical task in industrial manufacturing is machining, and the tools used for this are often worn down by use. Therefore, condition monitoring plays an important role in the manufacturing field. Here, machining tools are understood as all tools that can be used to machine materials, such as milling cutters, drills, saws, etc.

[0003] Generally speaking, the goal is to design tools to optimize usage time and thereby reduce costs.

[0004] Finding the optimal replacement point requires determining tool wear, ideally quantitatively, a time-consuming and labor-intensive task. On one hand, tools should not be replaced prematurely (preventatively), as they may still function with the desired quality for some time. Premature replacement is not only uneconomical and incurring higher costs, but also detrimental to sustainability. On the other hand, tools should not be replaced too late, as many workpieces will then fail to meet adequate quality, and the resulting scrap will similarly severely impair the manufacturer's productivity and sustainability. Furthermore, there is the risk of tool breakage. This is associated with potential damage to the spindle, machine, and workpiece, and the resulting extended downtime due to maintenance.

[0005] Furthermore, many factors come into play when determining the wear of machining tools during the manufacturing process, such as the materials used for the workpiece and tools, cutting parameters, the coolant supply that may be required, the temperature reached during machining, etc., which greatly increase the complexity of predicting the wear of machining tools.

[0006] Changes in cutting parameters during cutting can lead to nonlinear wear, where most previous solutions have failed. Furthermore, monitoring tools that are not brand new and show signs of use is difficult because their past wear processes are often unknown.

[0007] It is known that existing wear at the machining tool can be determined from microscopic records, particularly those of the cutting or machined surfaces. This is, for example, in... Figure 1 The images above are shown. Top views of drills (or end mills) 11 and 13 are shown respectively, along with microscopic evaluations of the images, particularly the cutting edges, shown in 12 and 14. The case below shows a new machining tool (i.e., at time point t0, where no wear is yet detected (wear rate 0%)). Wear is given here as a percentage, i.e., the share of the affected area relative to the total area of ​​the drill tip. xThe machining tool, that is, after a pre-specified period of use. Here, wear can be identified at the cutting edge 111, where the wear rate is 90%. This is also highlighted 121 and evaluated in the microscopic evaluation. For this type of inspection, the cutting process must be interrupted, and in the worst case, the tool must be removed from the machining machine, which results in prolonged downtime, so this is usually not done during continuous operation.

[0008] For monitoring purposes, records from the quality assurance of the build-up produced by the tooling process can also be used. Possible standards here include, for example, the condition of the cutting edge (precise, clear, or burred) or the condition of the milled surface (smooth or rough). Of course, different adverse results can have different causes and do not necessarily point to tool wear.

[0009] Using a camera during cutting is associated with poor image quality, for example, due to the use of a coolant that makes unobstructed recording difficult or even impossible. The fundamental advantage of this method lies in its quantifiability, as the wear of the tool can be determined pixel-perfectly. However, when this is done manually, it implies a significant amount of expense and work.

[0010] It is also known that tool wear is assessed based on process signals, typically by setting fixed thresholds for the signals. Examples of such signals include torque, spindle current that generates torque, tool life, usage time, or even the cutting path.

[0011] Once the set threshold is exceeded, a request to change the tool is generated and displayed.

[0012] In mass production, it's also possible to record a tool fingerprint during workpiece manufacturing with new tools, and derive a threshold from this fingerprint, for example. The drawback is that a new fingerprint must be recorded for each newly manufactured workpiece (e.g., with a new geometry). Furthermore, changing cutting parameters is not always feasible without issues. Moreover, a new threshold must be derived each time because the manufacturing process depends on many factors, as mentioned above.

[0013] Another possibility for determining tool wear is using external sensors, such as force and / or vibration sensors. However, this solution also has several drawbacks: sensors and evaluation units cannot be installed in every tooling machine according to the machine's preset geometry / layout. Furthermore, this solution is unsustainable as it uses only additional electronics for this purpose. Moreover, it is important to mention that the correlation between sensor readings and wear must be known.

[0014] Applications that can be implemented using known process signals for trend monitoring are also limited. If the process parameters used, such as cutting speed or feed rate, decrease, the amplitude of the current or torque also decreases. If we now assume that higher wear results in higher torque current, then an unclear record of wear will be obtained when cutting parameters decrease. For partially worn tools, as mentioned above, trend monitoring also presents difficulties because the state of the drill bit is initially unknown.

[0015] Another possibility is to use lifespan calculations for wear monitoring. However, these calculations are often discrete. Once there are deviations in the properties of the materials used (workpiece and material) or in the process parameters, the calculation becomes inaccurate. Summary of the Invention

[0016] Therefore, the object of the present invention is to describe a solution for predicting the wear of machining tools, which can be used in multiple ways, requires no preparation time, makes the most accurate predictions possible, does not have the aforementioned disadvantages, and provides a solution that is as sustainable as possible.

[0017] This objective is achieved by a method according to the features of claim 1 of the independent patent.

[0018] Furthermore, this objective is achieved by a computer program product according to the features of claim 10.

[0019] This objective is also achieved by means of a device according to the features of claim 11 of the independent patent.

[0020] Other embodiments of the present invention are covered by the dependent claims.

[0021] In the solution proposed below, no microscopic recording or measurement points of the tool or workpiece are required during the cutting process. The proposed interpolation method derives the wear at each time point from the recording of the signal, which is then used to train the neural network.

[0022] The end of tool life is not determined preventively (by calculation), but rather predictively, taking into account the actual condition of the tool. Therefore, this solution offers advantages over life calculation. Attached Figure Description

[0023] The invention is also illustrated below with reference to the accompanying drawings. Herein lies: Figure 1 This shows a microscopic record of new and worn machining tools. Figure 2 This shows the current flow during pretreatment. Figure 3 The current flow is shown with wear values ​​at two time points. Figure 4The current flow is shown by the current integral characteristic curve. Figure 5 The current integral characteristic curve is shown. Figure 6 This demonstrates the augmentation of data during training. Figure 7 This demonstrates training a neural network using sequential 4-tuples. Figure 8 This demonstrates the training of a neural network. Figure 9 This demonstrates inference using a trained neural network, and Figure 10 This provides an overview of the application of this method in industrial facilities. Detailed Implementation

[0024] Figure 10 This provides an overview of how the methods described below can be put into use.

[0025] Machine tools 2, such as drilling machines, milling machines, or similar machines, are assumed and operated in industrial facility 1. This equipment and machine are controlled here, and signals are measured and parameters are determined. These signals and parameters can be used to train a neural network 5. Here, the method according to the invention now helps to: predict the wear degree 6 of tools used in the machine tools with the aid of the NN, thereby providing support for the maintenance and upkeep of the facility and for maintaining the quality of workpieces machined with the tools.

[0026] The following prerequisites must be met in order to apply the proposed method: - During the training phase of an artificial intelligence model, process signals are required, and wear is derived based on these process signals.

[0027] like Figure 2 As shown, our embodiment uses a torque-forming current, or Current, which is continuously recorded throughout the entire lifespan of the machining tool. Recording begins at time point t0 and ends at t... T The end, where the recorded end is not defined by the breakage of the machining tool, but can vary. Then, the initial wear v0 (e.g., 0%) and the final wear v must be derived. T (For example, 90%). This can be done by briefly measuring it before a new drill bit is clamped into the spindle. The wear derived from this is v0. After the drill bit wears out and is replaced, v is then measured finally. T This will be combined later. Figure 5 Let me elaborate further.

[0028] This solution is based on time-dependent modeling of wear between two wear measurement points, taken at the beginning and end of the tool's lifecycle. The timing of the measurement is crucial as it should not impair the cutting process. These measurement points can be recorded using any measurement system. As an example (see...),... Figure 1 As described above, a microscopic record can be used to provide a precise wear measure v. This wear measure v can be a physical quantity (such as the width of the worn surface) or a percentage, as described above.

[0029] Measurements at only two points in time are feasible in manufacturing.

[0030] In addition to process signals and information about wear, parameters are also needed, such as tool characteristics (material of the machining tool, coolant used, drive method, motor power) and workpiece characteristics (material, etc.). These must also be available from the beginning to the end of the recording.

[0031] Wear-tag interpolation

[0032] Since there are only two wear measurement points per life cycle (v0 and v) T Next, time-related wear is modeled. In the first step, the signal from which wear is to be extracted must be preprocessed. The signal depicting wear is, for example, the spindle current i that forms the torque, as this current essentially reflects the force that the spindle must exert. The higher the wear, the more work the spindle must do, and the higher the current. However, since this only becomes visible at the moment of cutting, these segments of the signal must be detected. This can be done, for example, through a supervised neural network (NN, such as RNN or CNN) or through other, such as rule-based schemes. No-load current, no-cut, spindle acceleration, etc., must be removed, and only the actual cutting current is allowed to be considered, because only this actual cutting current reflects wear. After detecting the cutting signal, the cutting signal must be concatenated into a single signal. If the cutting parameters have not changed during the lifetime, the amplitude of the cutting signal can be schematically represented as, for example, looking like Figure 2 That's it. Plot the time progression on the X-axis (Time) and the measured current on the Y-axis (Current). Curve 20 shows the applied current, for example, at time point t. x ,21, the current is i tx ,twenty two.

[0033] However, in cases such as reducing the cutting speed during cutting, the current flow will also fluctuate, see [reference needed]. Figure 3 Example in the text. Here, at time point t x1 and t x2The detection of identical current values ​​at points 31 and 32 leads to ambiguity. This applies when two different CNC subroutines with different cutting parameters are called alternately. Therefore, this example illustrates that in a favorable design, if the cutting parameters change during monitoring, the current should not be used directly to read wear.

[0034] The two problems listed (one being the change in cutting parameters, and the other being the prediction of partial tool wear) can be solved by introducing a current integral characteristic curve (red). This characteristic curve is defined in the embodiment as follows:

[0035] exist Figure 4 In China, the following was adopted: Figure 3 An example with characteristic curve 30 is provided, supplemented by current integral characteristic curve 40. Thus, the wear degree can be derived at any point in time independently of the cutting parameters through the accumulation of current. Current curve 30 for current value i tx It also has the ability to handle multiple time points t x1 t x2 The ambiguous relationship, while the new characteristic curve 40 is unambiguous here, and has a different (higher) current value at each time point.

[0036] To derive wear and tear during the process, the neural network (NN) is trained in the next step. The labels required for training the NN (e.g., a recurrent neural network, RNN) are derived in the manner described above. This characteristic curve is only needed during NN training and is no longer required in the production phase.

[0037] The training of the neural network NN is 80%. Figure 8 The diagram illustrates this. As input to the neural network NN, a value set S, 81 is used, which contains the signal at a specific time point t and j previous signals preceding time point t, "looking back". Time point t must be greater than or equal to 0 and also less than or equal to T.

[0038] The number of sequences used in training can vary. However, this will affect inference. If, for example, four sequences are used in training, the time required for KI to predict the wear of used tools is determined. For example, in the case of a drill bit, four sequences with a sampling rate X would be needed to achieve the desired accuracy, and thus in the case of a milling cutter, perhaps six sequences with a sampling rate X might be required.

[0039]

[0040] Parameter P,82 contains all the information, such as cutting parameters and the characteristics of the tool and workpiece. It can contain one or more signals.

[0041]

[0042] exist Figure 7 and Figure 8 The diagram illustrates how, during training, different tuples {s} consisting of signals S and parameters P are processed. j p j The input 61 is provided to the neural network for training; in this specific example, it is a 3-tuple. The bar chart 60 shows the progression of tool wear over time, from 0% at time t=0 to the maximum possible (or measured) wear v at time T. t For example, 90%. The parameter tuple is selected from this time period, and the width of the window (i.e., the spacing between measurement points) is a variable. The neural network 80 predicts wear at time point t. 62, 83. The label v is known from the current integral characteristic curve. t 85 and predicted wear Together, these are used to optimize the network. The loss L, 84, is calculated from this. The loss is a so-called loss function that indicates how much more the neural network needs to "miss" to learn enough. The goal of training is to minimize the loss, and ideally, L=0 at the end of training.

[0043] Figure 6 and Figure 7 This shows how data is "fed" into the network, and Figure 8 Show training, Figure 9 The reasoning is shown.

[0044] Figure 8 The proposed training mechanism is shown: it is enhanced by randomly generating signals S, 81 and parameters P, 82, i.e., S and P have different j and t.

[0045] An example of a neural network's training mechanism can be seen as follows: Using the entire signal from t=0 to t=T (divided into X sequences). This signal has 100 measurement points and is divided into 20 sequences. Each sequence then has 5 measurement points. For training, 4 sequences are used. This results in 20 / 4 = 5 iterations.

[0046] Alternatively, it can enhance the data. Therefore, this can be viewed as follows: First traversal: Sequence 1, 2, 3, 4 Second traversal: Sequence 2, 3, 4, 5 Third traversal: Sequence 3, 4, 5, 6 The 4th traversal: sequence 4, 5, 6, 7, etc.

[0047] This can also be done randomly. This increases the number of iterations and provides more training material for KI.

[0048] The number of sequences can also vary. If the 100 measurement points are divided into 30 sequences, this can be done by stacking sequences at the edges. If the first sequence contains measurement points 1, 2, 3, and 4, then the second sequence starts from measurement point 3. By allowing the length of the sequences and the number of sequences in training to be changed randomly, KI becomes more robust to the massive amounts of data generated by augmentation.

[0049] The required labels describing the wear condition (e.g., in percentage) are calculated using a current integral characteristic curve. Thus, the neural network 90 utilizes the tuple {s} j p j} to {s t p t Predict the wear and tear of used tools. 91.

[0050] Production stage (reasoning)

[0051] When the start time of recording signals and measurements does not begin with the use of a new machining tool, this presents a problem for wear modeling. In a specific example, this means that accurate wear prediction during the production phase is more difficult because the past of the drill bit used is unknown.

[0052] Inference is already known and represents a productive use of neural networks—it computes the expected result, such as a score, segmentation, or classification, or calculates wear and tear in a pre-set example. Inference infers results for new objects of the same class based on the adaptation of individuals trained previously. Successful training allows inference to always arrive at the correct conclusion for that new object.

[0053] During inference, the trained model is applied (see...). Figure 9 Wear and tear on new, unused tools. It can make correct predictions from the beginning because the neural network knows that v is also equal to 0 at t=0. In the case of tools that are already partially worn, it takes some time until the network recognizes the specific pattern and can make correct predictions.

[0054] To monitor quality, it is possible to supplement this with quality assurance records from production of components machined by tools. Possible criteria, as described above, include the condition of the cutting edge (precise, clear, or burr-free) or the condition of the milled surface (smooth or rough).

[0055] By using a neural network trained in this way, it becomes possible to predict at what time t the tool under consideration will reach the point where it needs to be replaced.

[0056] The method described above has numerous advantages over known prior art: This enables wear prediction that is not constrained by batch production. For example, it does not use "fingerprints" as described in existing technologies. Nor does it require a fixed threshold to determine the tool's state, allowing the solution to be used for sequentially machining different workpieces. Similarly, cutting parameters can be changed during cutting, and the model can be adapted to those parameters.

[0057] Furthermore, no additional external sensors are required. Wear prediction is based solely on already available process signals and parameters, making this solution more sustainable overall than known solutions.

[0058] During the production phase, computing units, such as inter-process communication (IPC) / edge devices, are needed to enable the application of models. However, multiple applications can also run in parallel on a single device.

[0059] This solution also offers advantages over trend monitoring of the signal. Changes in cutting parameters during monitoring become feasible via the current integral characteristic curve. Wear prediction for partially used tools also becomes feasible through enhancements during training.

Claims

1. A computer-implemented method for predicting the wear of a machining tool (11, 12) used for machining materials in industrial manufacturing, the method comprising the following steps: - The wear degree (v0) of the machining tool is obtained at the first time point (t0). - at a second point in time (t T ) deriving a degree of wear (v T ) of the machining tool - Time-dependent wear was modeled by measuring the signals (20, 30) depicting wear and plotting them as time-dependent characteristic curves (30, 40). - The value set (S) is generated from the modeled characteristic curves (30, 40) depicting the wear signal, and - The neural network (NN) is trained by inputting the set of values ​​(S) and the set of parameters (P) relating to the characteristics of the processing tool or the material to be processed. - deriving, by means of said neural network, a time point of maximum wear (v t ) at which a request for replacing said machining tool is made.

2. The method for predicting the wear of machining tools (11, 12) according to claim 1, characterized in that, The wear degree was measured using microscopic recording with the aid of the machining tools (11, 12).

3. The method for predicting the wear of machining tools (11, 12) according to claim 1 or 2, characterized in that, The wear rate (v) is given as a percentage.

4. The method for predicting the wear of machining tools (11, 12) according to claim 1 or 2, characterized in that, The wear degree (v) is given as a physical quantity, and in particular, the wear degree is given as the shape, length or area of ​​the wear surface (111, 121) of the machining tool.

5. The method for predicting the wear of machining tools (11, 12) according to any one of the preceding claims, characterized in that, The signal describing wear is the driving current (20, 30) of the forming torque of the machining tool (11).

6. The method for predicting the wear of machining tools (11, 12) according to any one of the preceding claims, characterized in that, Only the actual cutting current utilizes the signal describing wear.

7. The method for predicting the wear of machining tools (11, 12) according to any one of the preceding claims, characterized in that, The wear-depicting signal (30) is corrected using the current integral characteristic curve (40), particularly based on... The correction is performed by accumulating the signal (30) depicting wear in the current integral characteristic curve.

8. The method for predicting the wear of machining tools (11, 12) according to any one of the preceding claims, characterized in that, Predictions are made by applying reasoning when the machining tools already in use have pre-existing wear marks.

9. The method for predicting the wear of machining tools (11, 12) according to any one of the preceding claims, characterized in that, Records of workpieces processed by the said machining tool are used for monitoring, particularly the construction of cutting edges or milled surfaces.

10. A computer program product, adapted and configured to perform the method according to any one of the preceding claims.

11. An apparatus for predicting the wear of a machining tool (11, 12) used for processing materials in industrial manufacturing, the apparatus being adapted and configured to train a neural network (NN, 90) using data from time-dependent characteristic curves (30, 40). - The characteristic curve is derived from a time-related wear model, calculated from measurements of the signal (20, 30) depicting wear: - The first wear degree (v0) of the machining tool at the first time point (t0). - a second wear degree (v T ) of the machining tool at a second point in time (t T ) - The set of values ​​(S, 81) of the signal depicting wear, derived from the modeled characteristic curves (30, 40), and - The neural network (NN, 90) is trained using data from the value set (S, 81) and other values ​​from the parameter set (P, 82) relating to the characteristics of the processing tool (11, 12) or the material to be processed. and the device is adapted and arranged for implementing the trained neural network (NN, 90) to derive, by the neural network (NN, 90), a time point of maximum wear (v t ) of the component (10). The device is also adapted and configured to output a time point for a request to replace the machining tools (11, 12).

12. The apparatus for predicting the wear of machining tools (11, 12) according to claim 11, characterized in that, The wear degree is obtained by microscopic recording of the machining tools (11, 12).

13. The apparatus for predicting the wear of machining tools (11, 12) according to claim 11 or 12, characterized in that, The wear rate (v) is given as a percentage.

14. The apparatus for predicting the wear of machining tools (11, 12) according to any one of claims 11 to 13, characterized in that, The wear degree (v) is given as a physical quantity, and in particular, the wear degree is given as the shape, length or area of ​​the wear surface (111, 121) of the machining tool.

15. The apparatus for predicting the wear of machining tools (11, 12) according to any one of claims 11 to 14, characterized in that, The signal describing wear is the driving current (20, 30) of the forming torque of the machining tool (11).

16. The apparatus for predicting the wear of machining tools (11, 12) according to any one of claims 11 to 15, characterized in that, Only the actual cutting current utilizes the signal describing wear.

17. The apparatus for predicting the wear of machining tools (11, 12) according to any one of claims 11 to 16, characterized in that, The wear-depicting signal (30) is corrected using the current integral characteristic curve (40), particularly based on... The correction is performed by accumulating the signal (30) depicting wear in the current integral characteristic curve.

18. The apparatus for predicting the wear of machining tools (11, 12) according to any one of claims 11 to 17, characterized in that, Predictions are made by applying reasoning when the machining tools already in use have pre-existing wear marks.

19. The apparatus for predicting the wear of machining tools (11, 12) according to any one of claims 11 to 18, characterized in that, Records of workpieces processed by the said machining tool are used for monitoring, particularly the construction of cutting edges or milled surfaces.

Images