Artificial intelligence in discrete manufacturing

Technical Field

[0001] The invention relates to the products and process as per the first portion of the independent claims.

Background Art

[0002] In CNC-Machining and robotics, numerical control (NC) or computerized numerical control (CNC) refers to the automated control of machine tools - for instance for, drilling, boring, turning, laser, 3D printing or robotic applications such as welding, handling or painting by means of a computer. A CNC machine alters a blank piece of material such as metal, plastic, wood, ceramic, or any composite thereof to meet precise specifications by following programmed instructions and without a manual operator.

[0003] PTL1 discloses a closed-loop control process for a machining tool based on a numerical control program.

[0004] The CNC machining process as per Figure 1 (prior art) will now be

explained.

[0005] First, the NC program (1) is generated offline by an NC programming

system as described, for example, in NPL1.

[0006] An essential software component of CNC controller commonly referred to as the interpolator (2) computes the most important commanded points for the programmed tool path. Another software task of the controller is known as the position controller and, based on the path, periodically calculates commanded axes positions, velocity and acceleration along each of the axes. The generated commanded position is transmitted to a drive or current controller which periodically and precisely computes the required electrical current. Hence, all commanded values are reliably and periodically updated periodically at varying time intervals. The interpolator task of today’s CNC-controller is being executed every 4 msec/250Hz, the position task at 1 msec/1.000 Hz, the drives control task works at a frequency of 50 microsec/2.000 Hz.

[0007] Once the commanded values described above have been calculated and provided to the drives as input, the machine is set in motion while the controller attempts to apply the commanded values such that the corresponding real values follow as closely as possible. Whenever the controller detects deviation (lag) of process variables from its commanded value, it attempts to eliminate that difference by various control algorithms. To perform such control operation, each CNC machine is equipped with position sensors that constantly measure the current position of each axis. From these sensor values (3), further variables such as speed,

acceleration, or jerk are being deduced. Furthermore, the drives provide the real current values and current acceleration for the execution of the drives control loop.

[0008] After the CNC machine has finished the machining of a workpiece,

measurements (4) are typically taken by means of conventional coordinate measuring machines (CMM) to inspect the piece’s geometrical and surface quality. The measurements can be defined in terms of tolerance, such as may apply to parallelism, perpendicularity, angularity, position, concentricity, or circularity. In typical manufacturing drawings,

corresponding symbols are used to specify tolerances such as

+3pm/-5pm. These measurements (4) are typically documented in reports that, besides defining tolerance and other pertinent quality criteria, quantify precisely any deviations thereof. If during measuring by a CMM-machine values are being detected out of the defined tolerance the tested workpiece is usually rejected.

Summary of invention

[0009] By applying artificial intelligence techniques such as neural networks or machine learning, it is possible to automatically generate a digital model that establishes the relationships between the input data (configuration data, NC program, commanded values) and output data such as the quality or the cycle time of the produced workpiece. [0010] With these models, it is possible to find out which input parameters are responsible, for example, proportionally for resulting quality or cycle time times. Likewise, it is now possible to set the target in the other direction, such as, for example, the workpiece quality, and to manipulate the input parameters in such a way that the desired goals are realized

Technical Problem

[0011] The quality of a workpiece can only be measured once it has been

machined. Since all transient commanded and real values are constantly deleted by the controller there is no data or correlation between the work piece quality and the commanded and real data available. If a quality problem is being measured on the measuring machine, each parameter of the NC program thus needs to be varied in a time-consuming trial-and- error procedure. Dynamic behavior of interpolator, position controller, and drive may hence be considered a“black box”, the detailed functioning of these control modules remaining largely unclear.

[0012] Also, the NC program is static and does not change during series

production. Even if the behavior of the machines changes in the long term due to its aging, the NC program as such remains unchanged, resulting in a creeping degradation of workpiece quality. There is typically no adaptation to changing production conditions.

[0013] The real problem of conventional CNC machining is the fact that there is no cause-and-effect model between the input data and the machining result.

Solution to Problem

[0014] Since conventional methods are incapable of describing the cause-and- effect relationship between input data and output data, artificial intelligence techniques such as neural networks or machine learning techniques are used to automatically and self-generate a digital model. The availability of the automatically generated digital model opens completely new potential. This makes it possible to find out very quickly in the CNC machining process which input data have caused a machining result or also to define the machining results and to adjust the gate parameters backwards in such a way that the desired work jobs can be entered with certainty.

Brief description of drawings

[0015] Figure 1 (prior art) shows a conventional CNC machining process.

[0016] Figure 2 illustrates Data-Preparation and Pre-Processing

[0017] Figure 3 illustrates the model-generation process

[0018] Figure 4 illustrates the quality root cause analysis. It identifies the NC- program-parameters impacting the part tolerances.

[0019] Figure 5 illustrates the prediction of part quality based on the digitally

generated model.

[0020] Figure 6 illustrates the online adaption of NC-program-parameters in order to accomplish the required part quality.

[0021] Figure 7 illustrates the root cause analysis of all NC-parameters which influence the cycle time of the manufacturing process.

[0022] Figure 8 illustrates the online adaption of the relevant NC-program

parameters which influence the cycle time of the machining process.

[0023] Figure 9 illustrates the applied Artificial Intelligence methods.

[0024] Figure 10 illustrates the data-driven modelling

[0025] Figure 11 illustrates how Machine Learning generates the digital model.

[0026] Figure 12 illustrates how Machine Learning optimizes the input

parameters.

[0027] Figure 13 illustrates the methodological components of the Machine

Learning process.

[0028] Figure 14 illustrates two methodological types of root cause analytics.

[0029] Figure 15 illustrates two methodological types of machine learning. Description of embodiments

[0030] In Phase 1 , the input data used is being prepared and pre-processed for the subsequent Quality and Productivity Analytic applications. The input data include:

• Configuration data of machine, controllers and drives

• NC program

• Dynamically generated setpoint data

• sensor data

[0031] In phase 2, the machine learning model is generated and parameterized.

The input data (configuration data of the machine, controllers and drives, NC program, dynamically generated setpoint data, sensor data and the output data are correlated.) The output data are the quality data measured by a measuring machine such as parallelism, rectangularity, centricity, concentricity or circularity accuracy. The manufacturing tolerances specified in all production processes are quantitatively evaluated during the workpiece measurement and it is determined whether these tolerances are in the defined window or outside, i.e. whether reject parts were produced by tolerance violation.

[0032] The machine learning therefore establishes the correlation between the input data and the actual production tolerances. The resulting quality data of the workpieces can in one case be predicted from the input data, in the other case can be identified from the measured manufacturing tolerances, which input parameters for the result was responsible at what impact. Finally, with this model the input data can also online being adapted during machining in order to produce 100% good workpieces.

[0033] After the data-preparation and pre-processing phase is being completed the Machine Learning process is now“learning” the digital model which maps the input with the output-parameters in form of a fully automates (self-learning” process (Figure 3)

[0034] The skilled artisan will appreciate that, beyond neural networks, other supervised learning techniques can be applied without departing from the scope of the invention, such as may be based on support vector machines, linear or logistic regression, naive Bayes or other probabilistic classifiers, linear discriminant analysis, decision trees, the r-nearest neighbor algorithm, or similarity learning.

[0035] Use Case 1 : Quality Root Cause Analytic of NC-Program Parameters

(Figure 4)

[0036] Figure 4 illustrates an example where a given tolerance - here, in terms of circularity - was violated with a deviation of 15 pm, the affected contours of the workpiece (6) being circled in the drawing. As may further be gathered from the figure, the proposed method (10 - Figure 2) pinpoints among the parameters of the NC program (1) the interpolation method, commanded feed rate, and tool data as causing this circularity defect with at an impact ratio of 53 % to 19 % to 28 %. With this knowledge, the NC programmer will readily apprehend that the deviation may be reduced by adapting said parameters. The conventional approach does not have a root-cause model between input- and output variables and a very time- consuming trial and error process must be executed. To reach an optimum is practically impossible.

[0037] Use Case 2: Online Model-based Quality Prediction (Figure 5)

[0038] Once a system as per the invention has learned the correlation model (8) and error criteria, it is possible - by what may be considered“virtual measuring” - to accurately predict the quality of a workpiece (6) being machined without taking physical measurements by a measuring machine. The physical measurement process is completely substituted by the described virtual measuring process. (4).

[0039] Use Case 3: Online Adaption of NC-Program Parameters in order to get 100% Part Quality Accomplished. (Figure 6)

[0040] Furthermore, the machine learning system may adapt parameters of the

NC program (1)“online” to minimize any quality breaches, striving for zero- defect production. (Figure 6). Since the machine learning system has generated a root cause model between all input- and dynamically generated commanded values, resulting real values and the resulting part quality the system can also be applied to achieve Zero-Defect

manufacturing results by adapting the NC-program parameters in a way that previously generated quality errors are now eliminated. [0041] Use Case 4: Online-Mode/Productivity-Root Cause Analytic by Artificial Intelligence (Figure 7)

[0042] The computational model of the process may be trained for optimization goals other than quality of the workpiece, for instance, the cycle time or net process time required by the tool for machining the workpiece. To this end, the network establishes any correlations between, for instance, parameters of the NC program and resulting cycle time. For example, the method identifies commanded feed rate, commanded acceleration, commanded jerk, commanded drives current and commanded power consumption of the spindle as influential at a ratio of 21 %, 27%, 19%, 21 % and 12%. Knowledge of these correlations will allow the NC programmer to adjust said parameters to eliminate wasteful expenditure of resources.

[0043] Use Case 5: Online-Mode/Productivity-Tuning (Optimization) by Artificial Intelligence (Figure 8)

[0044] Finally, once having explored the above correlation, an advanced

embodiment may fine-tune the identified parameters of the NC program online in order to minimize cycle time. In a first sub-use case, the system is trying to achieve the minimum cycle time. There might be strategic reasons to not use the minimum cycle time. For example, at maximum speeds the tools may age very fast so that the tool costs explode. If a strategic constraint like this is given the user may input to the machine learning system to reduce the cycle time for example by 45%, only down to 25% in order to combine the advantage of increased productivity on one hand and on the other hand to avoid exploding tool costs.

[0045] Applied Artificial Intelligence Methods (Figure 9)

[0046] Pre-processing the data is essential for good results. Correlation Analysis prunes the feature space by finding relevant or redundant features, while Autoencoders do the same by compressing and reconstructing input vectors. Principal Component Analysis also simplifies the data for the later stages by finding the features which best represent the data with minimal loss in information. Lastly, a Fast Fourier Transformation converts cyclic data (might be a subset of the original features) into its spectral representation which yields information about the main frequencies observed in the data. A change in frequency often is the result of an anomaly/error in real use cases and even small changes in the spectrum can be detected via the later stages.

[0047] The pre-processed data is then feed into a Machine Learning model,

which either clusters the data (k-Means), classifies (ANN with SoftMax output layer) or predicts (ANN or MARS) values on unseen data.

[0048] The output of the trained model when presented with new data can be used for root cause analysis, optimizing input parameters or predictive maintenance.

[0049] Machine Learning - Model Training (Figure 10)

[0050] The training step is using a model specific algorithm to learn the relation between the input data (sensors, parameters, ...) and the output (quality measurements, cycle times, error cases). The goal is a model which can produce accurate predictions about the output when presented with new input data, reducing for example the need for manual measurements.

[0051] Machine Learning - Model Training (Figure 11)

[0052] The process of model creation usually consists of three steps. A naive model tries to predict a known output value based on the corresponding input. By comparing the prediction with the actual value, it creates an error measure which it uses to update itself to minimize this error. This process is repeated for all datapoints in the training set or until the prediction is optimal.

[0053] Machine Learning - Optimization (Figure 12)

[0054] With the virtual model the system can also search for better combinations of input parameters given the environmental features and a metric to optimize (like no defects, shortest cycle time, most throughput).

[0055] Machine Learning - Regression Analysis (Figure 13)

[0056] Given the historic data, regression analysis tries to represent the

relationship between input and output by a set of mathematical functions with tunable parameters.

[0057] After training a model, data specific parameters get evaluated to find a better fit. This process gets iterated until the quality of the model is optimal. Training might be time consuming but deploying the model and analyzing new data points can be done with high frequency, usually only limited by the latency between machine data and model.

[0058] Machine Learning - Root Cause Analysis (Figure 14)

[0059] After the model training is complete, correlation analysis yields insight about the weight of impact each input dimension (sensors, parameters, ...) has on the output. Additionally, after analyzing a datapoint as anomalous, the feature relevance can be computed to find the root cause of the problem.

[0060] Machine Learning - Supervised and unsupervised learning (Figure 15)

[0061] There are two general types of machine learning models. Supervised

means that for each input vector the class label or output vector is known. This enables to learn the precise relation between in and output but often, the class labels or output measurements aren’t available. In that case, unsupervised learning can still cluster the data in meaningful ways and can yield more insights about the data including detecting anomalies by outlier detection or making classes of errors visible.

[0062] Machine Learning - Model Training

[0063] The process of model creation usually consists of three steps. A naive model tries to predict a known output value based on the corresponding input. By comparing the prediction with the actual value, it creates an error measure which it uses to update itself to minimize this error. This process is repeated for all datapoints in the training set or until the prediction is optimal.

[0064] The learned model can then be used to detect anomalies, classify errors or predict attributes. With new data, the model can also be updated to improve the precision or learn new classes.

[0065] Irregular behavior can also be traced back and attributed to individual features/sensors by calculating the distances between the model’s representation of normal behavior and current sensor values.

Industrial applicability

[0066] The invention is applicable, among others, throughout the CNC-controller and robot-based discrete manufacturing industry. Reference signs list

[0067]

1 Numerical control program

2 Dynamic interpolator, position and drive controller

3 Sensor values

4 Measurements

5 Tool

6 Workpiece

7 Training

8 Model

9 Productivity and cycle time

10 Method as claimed

1 1 Machine configuration data

Citation list

[0068] The following documents are cited hereinbefore.

Patent literature

[0069] PTL1 : GB 1019896 A (GEN ELECTRIC) 02.09.1966

Non-patent literature

[0070] NPL1 : SMID, Peter. CNC programming handbook: a comprehensive guide to practical CNC programming. Industrial Press Inc., 2003.