A data-driven machine spindle cutting torque monitoring method
By constructing a data-driven time series prediction model and using an LSTM network to monitor the servo signals of the machine tool spindle motor, the accuracy and cost issues of CNC machine tool spindle cutting torque monitoring are solved, making it suitable for online monitoring of various types of machine tools.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF AERONAUTICS & ASTRONAUTICS
- Filing Date
- 2023-12-26
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies make it difficult to accurately, cost-effectively, and without affecting the workspace monitor the spindle cutting torque on CNC machine tools, especially on large machine tools where the feed axis cutting force is difficult to predict accurately using servo monitoring signals.
By synchronously monitoring servo signals such as the output torque of the machine tool spindle motor and the spindle motion status, a data-driven time series single-step or sequence prediction model is constructed, and an LSTM network is used for online monitoring of the spindle cutting torque.
It achieves low-cost, non-disruptive monitoring of spindle cutting torque, is applicable to various types of machine tools, and has great potential for industrial applications.
Smart Images
Figure CN117733646B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of CNC machine tool technology, and in particular to a technology for monitoring the machining status of CNC machine tools by monitoring the spindle cutting torque using machine tool servo signals. Specifically, it is a data-driven method for monitoring the spindle cutting torque of a machine tool. Background Technology
[0002] CNC machine tools are widely used in advanced manufacturing fields such as aerospace, rail transportation, and shipbuilding, and are considered key technological equipment reflecting a nation's manufacturing capabilities. Monitoring the machining process on CNC machine tools is crucial for ensuring machining quality and equipment safety. Cutting force is the most direct and measurable physical quantity reflecting the machining state, and cutting force monitoring data can be applied to various scenarios such as tool wear / breakage monitoring, chatter detection, and adaptive control of cutting loads.
[0003] Traditional cutting force monitoring methods based on benchtop / rotary cutting force gauges are costly and may impact workspace, making them difficult to apply in practical industrial production processes. Servo monitoring signals and motion states of CNC machine tools reflect the force conditions of each axis, including cutting force information as an external disturbance. Therefore, theoretically, this information can be used for online cutting force prediction. Chinese Patent 2021107781399, "A Data-Driven Self-Monitoring Method for Cutting Force of CNC Machine Tools," authorizes an online prediction method for feed axis cutting force based on feed axis servo motor current and motion state. Chinese Patent 2022105841099, "A Dynamic Cutting Force Monitoring Method and System in the Cutting Process," discloses an online prediction method for feed axis cutting force based on the three-phase current and root mean square of the spindle motor. Both of these methods target feed axis cutting force, while spindle cutting torque is more conveniently applied in scenarios such as tool condition monitoring and chatter identification. Furthermore, for some large machine tools, the feed axis drive chain path is too long, resulting in significant signal energy loss, making it difficult to accurately predict the feed axis cutting force using servo monitoring signals. Therefore, there is an urgent need for a machine tool spindle cutting torque monitoring method that is applicable to various types of machine tools, accurate, low-cost, and does not affect the workspace. Summary of the Invention
[0004] The purpose of this invention is to address the problem of obtaining accurate monitoring parameters in existing machine tool spindle cutting force monitoring systems. This invention provides a data-driven method for monitoring machine tool spindle cutting torque. It synchronously monitors servo signals such as the output torque (or three-phase current) of the machine tool spindle motor and the spindle motion state, as well as the spindle cutting torque. Cutting experiments are conducted to obtain a spindle servo monitoring signal-cutting torque dataset. This dataset is then trained using a sequence-sequence supervised training method to train a data-driven time-series single-step prediction model. In practical applications, the corresponding servo monitoring signals are input into the prediction model for single-step real-time prediction or sequence prediction, thereby achieving online monitoring of the spindle cutting torque.
[0005] The technical solution of this invention is:
[0006] A data-driven method for monitoring the cutting torque of a machine tool spindle, characterized by comprising the following steps:
[0007] 1) Set process parameters such as rotational speed, depth of cut, width of cut, and feed per tooth. With the process parameters fixed, any cutting trajectory constitutes a cutting experiment. During the cutting process, synchronously monitor the output torque (or three-phase current) of the machine tool spindle motor, the servo signals such as the spindle motion status, and the spindle cutting torque to obtain a cutting experiment sample in the form of a time series. The spindle servo signal is the characteristic signal, and the spindle cutting torque is the label.
[0008] 2) Select different combinations of process parameters such as rotation speed, depth of cut, width of cut and feed per tooth to design cutting experiments. Use the method in step 1) to obtain a set of cutting experiment samples and construct a dataset of spindle servo monitoring signals and cutting torque containing different combinations of process parameters.
[0009] 3) The model adopts a data-driven time series prediction model, and trains the spindle servo monitoring signal-cutting torque single-step prediction model in a sequence-sequence supervision manner;
[0010] 4) When applying, input the spindle servo monitoring signal into the prediction model. The model starts from the zero initial state and performs single-step real-time prediction of the spindle cutting torque or sequence-sequence prediction of arbitrary length. The internal state of the model after the previous prediction should be used as the initial state of the next prediction.
[0011] The spindle cutting torque refers to the equivalent resistance torque acting on the tool along its rotation direction generated during the tool-workpiece cutting process. It is equal to the tangential cutting force on the tool multiplied by the tool radius and can be measured using, but is not limited to, a benchtop or rotary cutting force measuring instrument.
[0012] The cutting experiment described above begins when the actual spindle speed stabilizes at the commanded speed and the tool no longer contacts the material, and ends when the tool has completely cut out the material and the commanded spindle speed remains unchanged.
[0013] The output torque (or three-phase current) of the machine tool spindle motor is a necessary monitoring quantity, and output torque should be used preferentially when monitoring conditions permit. For synchronous spindle motors, the torque current can be monitored instead of the output torque (or three-phase current); alternatively, the root mean square (RMS) of the three-phase current can be used to further replace the torque current, but the RMS current needs to be assigned the corresponding positive or negative sign according to the motor's forward and reverse rotation.
[0014] The spindle motion state is an optional but not necessary monitoring quantity. If the monitoring conditions permit, the spindle speed can be added as an input, or both the spindle speed and the spindle angular acceleration can be added as inputs.
[0015] The spindle servo signal is preferably obtained directly from the CNC system. If it cannot be read from the CNC system, it can be obtained by using an external Hall sensor to monitor the three-phase current of the spindle motor or by connecting an external encoder signal (spindle motion status).
[0016] The cutting experiment was designed by selecting different combinations of process parameters such as rotational speed, depth of cut, width of cut, and feed per tooth. When selecting process parameters, sampling was carried out based on their actual application range and the degree of dispersion of their values.
[0017] The data-driven time series prediction model mentioned above is recommended, but not limited to, neural network-based time series prediction models, such as Long Short-Term Memory (LSTM) networks and Gated Recurrent Unit (GRU) networks.
[0018] The input of the spindle servo monitoring signal into the prediction model should begin after no cutting has occurred and the actual spindle speed has stabilized at the commanded speed.
[0019] In addition, the sampling frequencies of the spindle servo signal and the cutting torque should be consistent and should be at least twice the maximum cutting frequency in the predicted requirements.
[0020] In short, the feature of this invention is to monitor the output torque (or three-phase current) of the machine tool spindle motor and the spindle motion status signal. Then, a single-step prediction model of the servo monitoring signal to the spindle cutting torque is constructed using a data-driven time series model. Finally, the servo monitoring signal is input into the actual part cutting process. The single-step real-time prediction mode or the sequence-sequence prediction mode can be selected according to the application requirements of the actual processing scenario to realize online monitoring of the spindle cutting torque.
[0021] The beneficial effects of this invention are:
[0022] The method disclosed in this invention significantly reduces costs compared to traditional force sensing elements and does not affect the machine tool workspace, thus having great potential for widespread application in actual industrial production processes. Attached Figure Description
[0023] Figure 1 This is a schematic diagram of the LSTM network model of the present invention.
[0024] Figure 2 This is a cutting experiment scene diagram for the present invention.
[0025] Figure 3This is the cutting torque prediction effect of input mode 1 of the present invention.
[0026] Figure 4 This is the cutting torque prediction effect of input mode 2 of the present invention. Detailed Implementation
[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0028] like Figure 1-4 As shown.
[0029] A data-driven method for monitoring the cutting torque of a machine tool spindle includes the following steps:
[0030] 1) Set process parameters such as rotational speed, depth of cut, width of cut, and feed per tooth. With the process parameters fixed, any cutting trajectory constitutes a cutting experiment. During the cutting process, synchronously monitor the output torque (or three-phase current) of the machine tool spindle motor, the servo signals such as the spindle motion status, and the spindle cutting torque to obtain a cutting experiment sample in the form of a time series. The spindle servo signal is the characteristic signal, and the spindle cutting torque is the label.
[0031] 2) Select different combinations of process parameters such as rotation speed, depth of cut, width of cut and feed per tooth to design cutting experiments. Use the method in step 1) to obtain a set of cutting experiment samples and construct a dataset of spindle servo monitoring signals and cutting torque containing different combinations of process parameters.
[0032] 3) The model adopts a data-driven time series prediction model, and trains the spindle servo monitoring signal-cutting torque single-step prediction model in a sequence-sequence supervision manner;
[0033] 4) When applying, input the spindle servo monitoring signal into the prediction model. The model starts from the zero initial state and performs single-step real-time prediction of the spindle cutting torque or sequence-sequence prediction of arbitrary length. The internal state of the model after the previous prediction should be used as the initial state of the next prediction.
[0034] This implementation uses a classic neural network model suitable for time series forecasting—the Long Short-Term Memory (LSTM) network—to construct a nonlinear dynamic prediction model for the servo monitoring signal to the spindle cutting torque, and verifies it using cutting experiments. The LSTM network possesses adaptive long short-term memory capabilities and performs well in modeling time series data.
[0035] Figure 1 The diagram shows the structure of an LSTM network. An LSTM network has three gate structures, with the forget gate, input gate, and output gate arranged from left to right. The forget gate formula is as follows:
[0036] f t =σ(W if x t +b if +W hf h t-1 +b hf )
[0037] Where, x t Given the input at time t, h t-1 Let W be the hidden state at time t-1. if and W hf For the corresponding weight matrix, b if and b hf Let f be the corresponding bias vector, σ denote the sigmoid function, and f t This is the output of the forget gate.
[0038] The input gate formula is as follows:
[0039] i t =σ(W ii x t +b ii +W hi h t-1 +b hi )
[0040] Among them, W ii and W hi For the corresponding weight matrix, b ii and b hi For the corresponding bias vector, i t The output of the input gate.
[0041] To generate the cell state at time t, the candidate cell states at time t are first calculated.
[0042]
[0043] Among them, W ic and W hc For the corresponding weight matrix, b ic and b hc This is the corresponding bias vector. Further, after obtaining the forget gate output f... t Input gate output i t and candidate cell status Afterwards, the cell state ct at time t can be calculated using the following formula:
[0044]
[0045] Here, ⊙ represents element-wise multiplication, i.e., the Hadamard product.
[0046] The output gate formula is as follows:
[0047] o t =σ(W io x t +b io +W ho h t-1 +b ho )
[0048] Among them, W io and W ho For the corresponding weight matrix, b io and b ho For the corresponding bias vector, o t This is the output of the output gate. Finally, the hidden state h of the LSTM unit at time t... t From cell state c t and output gate output o t The decision was made jointly, and the calculation formula is as follows:
[0049] h t =o t ⊙tanh(c t )
[0050] Finally, the LSTM network outputs y t By h t After transformation by a linear fully connected layer, the following is obtained:
[0051] y t =W l h t +b l
[0052] Among them, W l For the corresponding weight matrix, b l This is the bias vector.
[0053] In this invention, there are six main modes for inputting x, as follows:
[0054] x = T e
[0055] x = [T] e ω] T
[0056]
[0057] x = [I u I v I w ] T
[0058] x = [I u I v I w ω]T
[0059]
[0060] Among them, I u I v I w Three-phase current of spindle servo motor, T e ω is the output torque of the spindle servo motor, and ω is the spindle speed. Main spindle acceleration, output is main spindle cutting torque T m When the spindle motor is a synchronous motor, T e The three-phase current can be replaced by the torque current; or the root mean square of the three-phase current can be used to further replace the torque current, but the root mean square current needs to be assigned the corresponding positive and negative values according to the forward and reverse rotation of the motor.
[0061] In the implementation case verification of this invention, the sampling frequency of both the spindle servo monitoring signal and the cutting torque signal was set to 4kHz. Based on the actual monitoring conditions, two input modes were set and compared:
[0062] x1=[T e ω] T
[0063] x2=[I u I v I w ω] T
[0064] The training set is configured with thirty combinations of depth of cut and feed rate, where the depth of cut a is... z The thicknesses are 2.5mm, 2.7mm, 2.9mm, 3.1mm, 3.3mm, and 3.5mm, with feed rates of 1500mm / min, 1625mm / min, 1750mm / min, 1875mm / min, and 2000mm / min, and a cutting width a. w The value is 25% of the tool diameter, the tool diameter D is 20mm, and the number of cutting edges N is... t For a depth of cut of 3, a straight path is machined using climb milling. The test set is also set as a permutation of depth of cut and feed rate, totaling fifteen groups, where the depth of cut a z The thicknesses are 2.6mm, 2.8mm, 3mm, 3.2mm, and 3.4mm, with feed rates of 1575mm / min, 1700mm / min, and 1900mm / min, respectively. All other parameters and processing methods are the same. The experimental scenario is as follows: Figure 2 As shown.
[0065] For input mode 1, the LSTM unit input is 2-dimensional; for input mode 2, the LSTM unit input is 4-dimensional. The LSTM unit output in both input modes is 1-dimensional, with a 512-dimensional hidden layer followed by a 512-dimensional to 1-dimensional linear fully connected layer. The initial state is set to a zero vector, and training is sequence-to-sequence, inputting a set of spindle servo monitoring signals and motion state signals for each set of parameters, with the label data being the spindle cutting torque sequence corresponding to that set of process parameters. Ultimately, the mean absolute error (MAO) for the training set of input mode 1 reaches 0.3567, and the MAO for the test set reaches 0.4754. For input mode 2, the MAO for the training set reaches 0.5963, and the MAO for the test set reaches 0.6737. Figure 3 and Figure 4 Prediction results for the same test set under two different input modes are presented. It can be seen that the method disclosed in this invention achieves good prediction results, and using motor torque as input yields better prediction results than using three-phase current.
[0066] The parts not covered in this invention are the same as those in the prior art and are implemented using existing technologies.
Claims
1. A data-driven method for monitoring the cutting torque of a machine tool spindle, characterized in that, Includes the following steps: 1) Set process parameters, including spindle speed, depth of cut, width of cut, and feed per tooth; with the process parameters fixed, any cutting trajectory constitutes a cutting experiment. During the cutting process, the spindle servo monitoring signal and the spindle cutting torque are monitored synchronously; the spindle servo monitoring signal includes the output torque or three-phase current signal of the machine tool spindle motor and the spindle motion status signal, resulting in a cutting experiment sample in the form of a time series, wherein the spindle servo monitoring signal is the characteristic signal and the spindle cutting torque is the label; 2) Select different combinations of process parameters such as rotation speed, depth of cut, width of cut and feed per tooth to design cutting experiments. Use the method in step 1) to obtain a set of cutting experiment samples and construct a dataset of spindle servo monitoring signals and cutting torque containing different combinations of process parameters. 3) The model adopts a data-driven time series prediction model, and trains the spindle servo monitoring signal-cutting torque single-step prediction model in a sequence-sequence supervision manner; 4) Input the spindle servo monitoring signal into the prediction model. The model starts from zero initial state and performs single-step real-time prediction of spindle cutting torque or sequence-sequence prediction of arbitrary length. The internal state of the model after the previous prediction should be used as the initial state of the next prediction.
2. The data-driven machine tool spindle cutting torque monitoring method according to claim 1, characterized in that: The spindle cutting torque refers to the equivalent resistance torque acting on the tool along its rotation direction generated during the tool-workpiece cutting process. It is equal to the tangential cutting force on the tool multiplied by the tool radius and can be measured using, but is not limited to, a benchtop or rotary cutting force measuring instrument.
3. The data-driven machine tool spindle cutting torque monitoring method according to claim 1, characterized in that: The cutting experiment described in step 1) begins when the actual spindle speed stabilizes at the commanded speed and the tool no longer contacts the material, and ends when the tool has completely cut out the material and the commanded spindle speed remains unchanged.
4. The data-driven machine tool spindle cutting torque monitoring method according to claim 1, characterized in that: Step 1) The output torque or three-phase current of the machine tool spindle motor is a necessary monitoring quantity. If the monitoring conditions permit, the output torque is used. If the spindle motor is a synchronous motor, its torque current is monitored instead of the output torque or three-phase current. Alternatively, the root mean square of the three-phase current can be used to further replace the torque current, but the root mean square current needs to be assigned the corresponding positive and negative values according to the forward and reverse rotation of the motor.
5. The data-driven machine tool spindle cutting torque monitoring method according to claim 1, characterized in that: Step 1) The spindle motion state described should be input by adding spindle speed as input, or by adding two motion states, spindle speed and spindle angular acceleration, as input, if monitoring conditions permit.
6. The data-driven machine tool spindle cutting torque monitoring method according to claim 1, characterized in that: The spindle servo monitoring signal mentioned in step 1) can be directly obtained from the CNC system. For cases where it cannot be read from the CNC system, an external Hall sensor is used to monitor the three-phase current of the spindle motor, and an encoder signal is used to monitor the spindle motion status.
7. The data-driven machine tool spindle cutting torque monitoring method according to claim 1, characterized in that: Step 2) describes the design of cutting experiments by selecting different combinations of process parameters such as rotational speed, depth of cut, width of cut, and feed per tooth. When selecting process parameters, sampling is performed based on their actual application range and the degree of dispersion of their values.
8. The data-driven machine tool spindle cutting torque monitoring method according to claim 1, characterized in that: Step 3) describes a data-driven time series prediction model, which employs a neural network-based time series prediction model, including a Long Short-Term Memory (LSTM) network.
9. The data-driven machine tool spindle cutting torque monitoring method according to claim 1, characterized in that: Step 4) The input of the spindle servo monitoring signal into the prediction model should begin after no cutting has occurred and the actual spindle speed has stabilized at the commanded speed.
10. The data-driven machine tool spindle cutting torque monitoring method according to claim 1, characterized in that: The sampling frequencies of the spindle servo monitoring signal and the cutting torque should be consistent, and should be at least twice the maximum cutting frequency in the predicted requirements; if monitoring conditions permit, they should be at least five times.