WO2011026638A1

WO2011026638A1 - Method for classifying a laser process and a laser material processing head using the same

Info

Publication number: WO2011026638A1
Application number: PCT/EP2010/005434
Authority: WO
Inventors: Lngo Stork Genannt Wersborg
Original assignee: Precitec Kg; Precitec Itm Gmbh
Priority date: 2009-09-04
Filing date: 2010-09-03
Publication date: 2011-03-10
Also published as: EP2473314A1

Abstract

The present invention relates to a method for classifying a laser process of a workpiece by means of a laser material processing head, comprising the steps of recording acoustic data caused by the laser process of the workpiece, transforming the acoustic data into a frequency domain by means of a wavelet decomposition or a windowed fourier transform to generate acoustic frequency features and classifying the acoustic frequency features on the basis of learned acoustic frequency features.

Description

Method for classifying a laser process and a laser material processing head using the same

Description

The present invention relates to a method for classifying a laser process and to a laser material processing head using the same, in particular to a cognitive approach for a robotic welding system learning how to weld from acoustic data. In industrial production flexibility becomes increasingly important to create complex products in many different variations. Coming from mass production the challenge is to realize mass customization. Thus, production units need to be able to calibrate to new processes in a minimum amount of time and labor costs. For laser material processing units this is in particular difficult, since it takes great efforts with many manual trials to set-up or recalibrate for new material processing tasks. A research aim is to use machine learning techniques to reduce these efforts and enable production systems to learn their job quicker, to realize if they are making a mistake, and for the best case to learn how to prevent them before they can happen. In production systems laser beam welding is chosen for high quality joining of materials. However, for industrial production these systems have to be set-up and calibrated manually with high efforts.

It is an object of the present invention to apply intelligent data processing resulting in a cognitive technical system which can learn how to weld, to speed up the configuring process, and reduce costs.

This object is solved by the method according to claim 1 and by a laser material processing head according to claim 14. Further advantages, refinements and embodiments of the in- vention are described in the respective sub-claims. Next to monitoring laser welding with cameras and optical sensors the present invention emphasizes the gain for monitoring with acoustic sensors and feature extraction. Using acoustic sensors the cognitive system is more insensible to strong optical radiation. Several combined methods such as wavelet analysis, fast Fourier transformation (FFT), and Princi- pie Component Analysis (PCA) are evaluated with sensor data from real experiments. Finally as machine learning, the results are classified with learned reference data to obtain reliable information for monitoring and possible closed-loop control.

When processing materials with laser light, strong radiations and process emissions occur. Different sensors have been proposed to monitor a laser beam welding process, such as photodiodes sensible for certain wavelengths and camera system observing a process before, within and after the welding process interaction zone, called pre-, in- and postprocess observing systems. According to the present invention, these sensor systems are used as input for the examined cognitive technical system. Furthermore, the existing industrial systems are extended with acoustical sensors for air and solid borne acoustics. This way a great amount of sensor data is gained during processing, enabling the characterization of the processes better than with just optical sensors. Especially, for materials with high reflectivity and relatively small melt pool sizes such as when processing Aluminum, acoustic sensors may support existing monitoring methods significantly. This enables the cognitive technical system to differentiate many different processes, if the difference is detected by one of the sensors used. Since the system can differentiate process characteristics it can also learn from a human expert, which are good or poor welds and to classify them accordingly. Laser beam welding is a well researched method with related work towards process observing systems and sensors. Approaches for laser material processing and using classifiers such as Artificial Neural Networks (ANN), Support Vector Machines (SVM) and Fuzzy Logic has been discussed such as the idea to have a self-learning system for laser beam welding. Acoustic sensors have already been investigated for laser beam welding. The ad- vantage of the cognitive approach of this invention is the combination of dimensionality reduction techniques and classifiers, which enables to gain more information from multiple sensors, especially acoustic sensors, within a reasonable time frame. Furthermore, this ap- proach might reduce the effort and costs of setting up or re-calibrate laser beam welding systems. This application aims towards a cognitive factory, which should enable a flexible way with intelligent or cognitive technical systems for industrial production of the future. To fulfill the above object, the present invention provides a method for classifying a laser process of a workpiece by means of a laser material processing head, comprising the steps of: recording acoustic data caused by the laser process of the workpiece, transforming the acoustic data into a frequency domain by means of a wavelet decomposition or a windowed fourier transform to generate acoustic frequency features, and classifying the acous- tic frequency features on the basis of learned acoustic frequency features.

Preferably, the acoustic data comprises solid-borne acoustic data or air-borne acoustic data.

Preferably, the acoustic frequency features are classified by means of a support vector classifier.

Preferably, the laser process is a laser welding process.

The method according to the present invention preferably further comprises the step of dimensionality reduction to generate acoustic frequency features having a reduced amount of non-relevant or correlating information.

Herein, the step of dimensionality reduction preferably comprises at least one of the methods selected from the group comprising Isometric feature mapping (Isomap), Principal Component Analysis (PCA), Locally Linear Embedding (LLE), or multidimensional scaling (MDS).

According to an advantageous embodiment of the present invention, only the first component obtained from dimensionality reduction is used to classify the laser process.

Preferably, the learned acoustic frequency features are acquired during an initial training, in which a human expert classifies a plurality of acoustic frequency features. Herein, classifying the acoustic frequency features preferably comprises the detection of a full penetration event through the workpiece.

In another embodiment of the present invention, classifying the acoustic frequency features comprises the classification into the process states of not enough laser power, full penetration achieved, optimal laser power, and to much laser power.

In still another embodiment of the present invention, classifying the acoustic frequency features comprises identifying the characteristics learned from previous workpieces when processing a new workpiece with different material properties.

The method of the present invention preferably further comprises the step of controlling processing parameters on the basis of the classification result in a closed-loop control.

According to a further embodiment, the method further comprises the recording of further sensor data to be fed into the classification process.

The object of the present invention is further solved by a laser material processing head, which comprises a control unit being adapted to perform the method of the present invention.

Preferably, the laser material processing head further comprises piezo sensors to be mounted on the workpiece for detecting solid borne-acoustic signals from the workpiece.

The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this application, illustrate embodiments) of the invention and together with the description serve to explain the principle of the invention. In the drawings:

FIG. 1 shows a wavelet packet decomposition for n = 3;

FIG. 2 is a logarithmic representation of computed FFT coefficients X_n for solid-borne acoustics before and after increase of laser power; FIG. 3 shows a normalized power spectral density for different spectral bands of solid- borne acoustics for a sudden increase of laser power at 3 seconds;

FIG. 4 shows a wavelet packet decomposition for different spectral bands of solid-borne acoustics for a sudden increase of laser power at 3 seconds;

FIG. 5 shows an one dimensional outcome of dimensionality reduction applied on video data acquired during a sudden increase of laser power;

FIG. 6 shows a welding seam for n = 3, a is top view and b is bottom view;

FIG. 7 shows an one dimensional outcome of dimensionality reduction applied on a frequency space representation of solid-borne acoustics acquired during a linear increase of laser power on workpieces of different thicknesses;

FIG. 8 shows a wavelet packet decomposition for two spectral bands applied on solid- borne acoustics acquired during a linear increase of laser power on workpieces of different thicknesses;

FIG. 9 shows one dimensional outcomes of dimensionality reduction applied on video date acquired on two workpieces during a continuous increase of laser power; and

FIG. 10 shows predicted classes for a training and a test workpiece, processed by a laser of linearly increasing power from 300 to 1200 W.

First of all, the techniques used to process the sensor data and the experimental setup are described. Then, two embodiments of the present invention are discussed. In the first embodiment, laser power variation is used to determine changes within the solid-borne acoustics, whereas the second embodiment aims to use solid-borne acoustics for monitoring purposes. Next, a classifier is trained using machine learning to allow an autonomous monitoring and quality estimation of the welding process, solely based on acoustic features within a cognitive framework.

According to the present invention, signal analysis based on Fast Fourier Transform (FFT) and Wavelet signal transformations are used to preprocess recorded data. The amount of data is hereby not reduced but offers the later possibility to analyze the signal's power den- sity in different frequency ranges. Next, application of dimensionality reduction allows to determine which frequencies are related to changes within the process. Changes might be caused by variation of the applied laser power or by workpiece properties such as its thickness or chemical constitution.

In order to compare the results obtained from analysis of the solid-borne acoustics, video data is reduced in its dimensionality and used as reference signal. The method for classifying a laser process of a workpiece by means of a laser material processing head will be described in the following.

A. Windowed Fourier transform

As the aim of analysis of solid-borne acoustics is to reveal specific events or process states limited in time, the signal analysis in frequency domain requires a resolution in time. The windowed Fourier transform (WFT) splits a signal y(t) into sections of specified length 2n. Each of these sections are then transformed into frequency domain consisting of Fourier coefficients X_n.

B. Wavelet decomposition

Wavelets are known to have some advantages over Fourier transformations. As they are able to produce sharp peaks, they can be applied to non-periodic and non-stationary signals like those produced during laser processing.

The Wavelet packet decomposition in particular seems well suited for this environment. An orthogonal set of high- and low-pass filters is computed for a chosen mother-wavelet. As visualized in Figure 1 , the signal is processed in a number 1 of decomposition steps. For each step, both a high- and a low-pass filter are applied to every portion of the signal resulting from previous decomposition steps. As a result, we decompose the signal into n = 2¹ frequency bands. For our experiments we used a db4 mother-wavelet. Daubechies wavelets are a set of orthogonal wavelets with a maximal number of vanishing moments, which means that they are particularly well suited to encounter signal discontinuities. We choose a level 6 de- composition providing n = 64 logarithmically distributed frequency bands. For dimension reduction, we do not save all of those coefficients, but use only the mean energy per frequency band over a set of samples for further calculations. C. Dimensionality reduction

Dimensionality reduction techniques play a crucial role in the development and implementation of technical cognition. High-dimensional data sets represent manifolds of the perceived environment or process. Still, these manifolds contain large parts of non relevant or correlating information. The dimensionality reduction tries to reduce these useless information, thus also allowing representation of the manifold with reduced dimensionality.

In the past, several approaches like Isometric feature mapping (Isomap), Principal Component Analysis (PCA), Locally Linear Embedding (LLE), or multidimensional scaling (MDS) have been shown to be useful methods. They differ in their accuracy, speed, and ability to handle with non-linear dimensionality reduction. According to the present invention, the combination of PCA and Isomap is a powerful combination. It can be used as a pre-whitening tool, thus already reducing the manifolds in dimensionality. The resulting dimensions still revealed high correlations, but could not be removed by means of the PCA due to their non-linear relationship. Thus, the subsequent use of the Isomap technique successfully removes these correlations.

The results shown in this application feature only the first component obtained from dimensionality reduction, as it is sufficient to detect laser power induced changes within the solid-borne acoustics and video signal.

D. Classification

Classifiers are used to recognize patterns within a multidimensional feature space and hence require an initial training. Once the pattern to be recognized have been taught, the classifier outputs probabilities or predictions about the fed, unclassified features. Tradi- tional approaches aim to measure the geometric distance between these features and training features, but lack in reliability if training features of different classes are convoluted.

The Support Vector Classifier (SVC) allows separation of these classes by mapping the feature space into a higher dimensional space in which the separation can be performed by means of linear separating planes. The complexity of the laser welding dynamics leads to the assumption that non-linear emissions may occur, hence the use of the Support Vector Classifier is preferred according to the present invention. Experimental setup

In order to record solid-borne acoustics of the welding process, two piezo-sensors were mounted on the workpiece. Their signals were gained and sampled at 2 MHz by a data acquisition board configured and driven by Matlab. An additional co-axial CMOS video camera was used to acquire in-process videos at 1000 frames per second.

The processing was performed within six to seven seconds on stainless steel with sheets of a thickness of either 0.6 or 1.0 mm. Laser power variation in between 300 and 1200 W allowed to alter the process and thus to change its acoustical emissions.

Observing laser power variation with solid-borne acoustics

Acoustic emissions occur under stimulation of for e.g. mechanical influences. They are related to small displacements and are useful for classification tasks if they occur periodi- cally. Their properties mainly depend on the stimulation force and frequency and physical properties of the propagation medium. In case of solid-borne acoustics, they propagate through the processed material and are caused by external sources and the welding process itself. The bandwidth is in the range of MHz, thus solid-borne acoustics are mostly inaudible.

The use of solid-borne acoustics for monitoring purposes of laser welding processes is still unexploited. Noisy surroundings make it difficult to record data with high noise-to-signal ratio. Even more, conventional signal analysis in the frequency domain takes in account all frequencies if no band pass filter is used, thus irrelevant information may be processed.

Two different trials were performed to analyze the solid-borne acoustics of laser welding processes. At first, a workpiece with a thickness of 1.2 mm is processed with a laser power of 550 W on the first half, and 750 W on the second half. In consequence, signal variations occurring at the middle of the workpiece are caused by laser power variation and may reveal which solid-borne emissions are laser power dependent. A second trial analyzes how events depending upon the applied laser power can be detected. This example focuses on the event of full penetration which occurs for a specific laser power. The required laser power depends upon the processing velocity, joint geometry, laser beam, and workpiece properties. Step response of a 200 W laser power increase

The rapid change of process parameters allows to determine the step-response of a system. Here, the laser power was increased by 200 W after three seconds. The subsequent analysis of solid-borne acoustics should reveal how emissions depend upon the applied laser power. At first, a continuous laser power of 550 W was applied, allowing full penetration of the processed 1.2 mm workpiece. The sudden increase to 750 W led to an increase of absorbed energy and thus modified the resulting weld seam.

The six second long process resulted in 12.000.000 samples which were transformed in a first step with a 4096-point FFT into 2929 samples in frequency domain. Figure 2 depicts a logarithmic representation of the solid-borne acoustics in frequency domain for a time period of 0.5 seconds before and after the increase of laser power. One notes the decrease of density for frequencies above 500 kHz after the laser power step. A subsequent power spectral density analysis was performed on distinct spectral bands, aiming to reveal how changes in frequency domains perform in time domain. Therefore, the signal was split into 16 spectral bands with a bandwidth of 62.5 kHz each. Figure 3 shows the power spectral density (PSD) for the four frequency bands 0-62.5 kHz, 375- 437.5 kHz, 437.5-500 kHz, and 875-937.5 kHz. As expected, only higher frequencies are affected by the change of laser power applied three seconds after the process begun. The analysis of this process with the wavelet packet decomposition was performed with 1 = 4, resulting in 16 distinct values describing the relative power spectral density of the observed spectral band. Figure 4 shows its outcome for the same frequency bands as in the previous trial. Again, the change of laser power is only observable at higher frequencies, even though less distinguishable for frequencies around 500 kHz.

At last, the in-process video was processed in order to obtain independent components corresponding to variances within the pixels' values and thus allowing to obtain an independent feature related to the increase of laser power. It is concluded that laser power variation leads to changes in frequency domain. The solid- borne acoustics partially depend upon the applied laser power. In consequence, assuming a constant laser power, changes within the workpiece properties may lead to changes of emitted solid-borne acoustics. Thus, monitoring and control techniques can take advantage of the process' acoustics.

Linear increase of laser power

Having seen that both in solid-borne acoustic and video data changes of laser power are observable, it is now analyzed how events depending upon the applied laser power can be detected. Again, the analysis is based on frequency domain representation performed by FFT, wavelet packet decomposition, and dimensionality reduction performed on video data.

For this trial, workpieces of different thicknesses were processed for seven seconds. Dur- ing this time, the laser power was linearly increased from 300 to 1200 W, leading to a spontaneous full penetration event once the fed power was sufficient. Data transformation into the frequency domain was performed the same way as in the previous trial. Here, the outcome of the FFT was fed to a set of dimensionality reduction techniques, aiming to distinguish all of the relevant features related to the full penetration event from other emissions. The outcome for signals acquired from workpieces with a thickness of 0.6 and 1 mm is shown in Figure 7. Full penetration occurred after 2.4 seconds for the thinner workpiece and after 2.7 seconds for the thicker workpiece, reflecting the higher demand of fed power for increasing thickness. A weld seam analysis by an expert confirmed that the full penetration was achieved at these points. Hence, the change within the frequency domain of solid-borne acoustics can be useful to detect the full penetration event.

Application of the wavelet packet decomposition resulted in n = 4 signals, from which the two signals related to frequencies higher than 500 kHz revealed features similar to the outcome of the dimensionality reduction. Figure 8 shows the computed power spectral density for frequencies in the range of 500 to 750 kHz. Due to a low temporal resolution of 140 ms only, the exact time of full penetration cannot be measured in this plot. Still, the wavelet analysis required only application of one technique. As result, the wavelet package decomposition can be more powerful than the subsequent use of FFT and dimensionality reduction.

In order to compare the emissions and their processing of solid-borne acoustics, co-axially obtained video data of the process was reduced in dimensionality. Figure 9 depicts the obtained feature amplitudes. Here, the event of full penetration is less distinguishable than in the acoustics. It is therefore suggested to use solid-borne acoustics to monitor the event of full penetration.

Classification of process states based on acoustic features

In this section, the Support Vector Machine (SVM) is used as a classifier. Based on fea- tures obtained from analysis of the solid-borne acoustic, it should determine the process state and hereby allow to evaluate the process quality. Moreover, the subsequent use of computed probability estimates shall allow to control the applied laser power in order to reach the best achievable quality.

During the previous trials, the workpieces processed by a laser of increasing power were subject to different process states: not enough laser power (state 1), full penetration achieved (state 2), optimal laser power (state 3), and too much laser power (state 4). These four states are used to create four patterns to be recognized by means of the SVM.

The initial training consisted of features acquired during these process states, for which an expert assigned the corresponding labels listed above to the time series 0.3 - 2.1, 2.7 - 3.0, 3.3 - 3.9, and 4.9 - 6.1 seconds. The features were obtained by reducing the WFT of the signal in dimensionality and a subsequent out-of-sample extraction of all of the features. In total, ten features were used for classification tasks, compared to 4096 initial features. The outcome of the SVM training is a reduced set of features which are suitable in terms of reparability to other patterns, wherefrom originates the term of Support Vector.

A first test of the trained classifier is performed with the entirety of features from the training workpiece. The straight line of figure 10 depicts the predicted classes. All of the features are correctly assigned to the labels 1-4, apart from irregularities while the state chan- ges.

At next, sensor data of a similar process during which the same laser power gradient was applied is used to test the classifier. The predicted classes are shown by the dashed line in figure 10. Again, small classification errors appear at state transition, whereas the classifi- cation accuracy is close to 100% during the time-series used for the training. The classification results may not only be used for monitoring they can be applied to realize learning new materials and for closed-loop control. Learning new materials means that a classifier such as Support Vector Machines or Artificial Neuronal Networks (ANN) identifies the characteristics learned from previous workpieces when processing a new workpiece with different material properties. This may happen when changing a work load. The newly learned characteristics result in monitoring improvements for the new workload. For closed-loop control we propose to use the classification technique to choose the appropri- ate controller. Every classification class has its own PID controller based on the error minimization of the most prominent features and the Neuronal Network or the Support Vector Machine simply chooses the controller to be used at the moment. This technique enables both fast and very adaptive closed-loop control.

These results confirm the usability of features acquired from solid-borne acoustic sensors for classification purposes in laser welding processes. Still, workpiece characteristics should not change and ambient noise may not be altered. The present invention shows the usability of solid-borne acoustics in laser welding process for monitoring and control purposes in intelligent or cognitive technical systems. Thus, techniques known from machine learning and feature extraction have been found, allowing to obtain meaningful characteristics which can be used to train the system not only to find features within the taught process, but also in similar welding processes.

Three signal processing techniques are used in the present invention to detect acoustical signals in laser beam welding processes as well as confirmation of their usability. The analysis in frequency domain allowed to confirm that the solid-borne acoustics depend on the applied laser power. For this purpose, tests were performed with the WFT, PSD analysis and wavelet packet distribution technique. All of these determined which frequencies are relevant to be monitored. Comparison with a feature obtained from a co-axial video confirmed the correct determination of laser power variation within the solid-borne acoustics.

An examination of signals obtained while the laser power was continuously increased, de- picted once more that influences of varying laser power and thus changing process states can be observed. All of the shown techniques correctly detected the full penetration event. The influence of the workpiece thickness was seen as delayed full penetration weld. In consequence, workpiece properties can be estimated by analysis of the emitted solid-borne acoustics if the applied laser power is known. The Support Vector Classifier revealed its usability in terms of classification of laser welding processes affected by changes of laser power. Once trained, the classifier can recognize taught patterns and hereby estimate the process quality. The development of intelligent monitoring and control systems needs to take in account the obtained features, which are specific to changes of laser power or processes states. Moreover, parameters characterizing the functionality of the WFT or wavelet decomposition have to become process independent in order to cope with the variety of existing welding processes.

Future work will concentrate on the use of wavelet decomposition used to feed self- learning classifiers. Moreover, a more intense analysis will be performed, aiming to build a model of the acoustical system of processed workpieces. Further, neural networks can be trained to adopt the laser power according to the recorded acoustical features and hereby achieved a closed-loop control.

Claims

1. A method for classifying a laser process of a workpiece by means of a laser material processing head, comprising the steps of:

- recording acoustic data caused by the laser process of the workpiece;

transforming the acoustic data into a frequency domain by means of a wavelet decomposition or a windowed fourier transform to generate acoustic frequency features; and classifying the acoustic frequency features on the basis of learned acoustic frequency features.

2. The method according to claim 1, wherein the acoustic data comprises solid-borne acoustic data or air-borne acoustic data.

3. The method according to claims 1 or 2, wherein the acoustic frequency features are classified by means of a support vector classifier.

4. The method according to one of the preceding claims, wherein the laser process is a laser welding process.

5. The method according to one of the preceding claims, further comprising the step of dimensionality reduction to generate acoustic frequency features having a reduced amount of non-relevant or correlating information.

6. The method according to claim 5, wherein the step of dimensionality reduction comprises at least one of the methods selected from the group comprising Isometric feature mapping (Isomap), Principal Component Analysis (PCA), Locally Linear Embedding (LLE), or multidimensional scaling (MDS).

7. The method according to claim 5 or 6, wherein only the first component obtained from dimensionality reduction is used to classify the laser process.

8. The method according to one of the preceding claims, wherein the learned acoustic frequency features are acquired during an initial training, in which a human expert classifies a plurality of acoustic frequency features.

9. The method according to one of the preceding claims, wherein classifying the acoustic frequency features comprises the detection of a full penetration event through the workpiece.

10. The method according to one of the claims 1 to 8, wherein classifying the acoustic frequency features comprises the classification into the process states of not enough laser power, full penetration achieved, optimal laser power, and to much laser power.

1 1. The method according to one of the claims 1 to 8, wherein classifying the acoustic frequency features comprises identifying the characteristics learned from previous work- pieces when processing a new workpiece with different material properties.

12. The method according to one of the preceding claims, further comprising the step of controlling processing parameters on the basis of the classification result in a closed- loop control.

13. The method according to one of the preceding claims, further comprising the recording of further sensor data to be fed into the classification process.

14. A laser material processing head, comprising a control unit being adapted to per- form a method as claimed in one of the preceding claims.

15. The laser material processing head, further comprising piezo sensors to be mounted on the workpiece for detecting solid borne-acoustic signals from the workpiece.