Classifications

• • G—PHYSICS
  • G01—MEASURING; TESTING
  • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
  • G01N29/00—Investigating or analysing materials by the use of ultrasonic, sonic or infrasonic waves; Visualisation of the interior of objects by transmitting ultrasonic or sonic waves through the object
  • G01N29/44—Processing the detected response signal, e.g. electronic circuits specially adapted therefor
  • G01N29/46—Processing the detected response signal, e.g. electronic circuits specially adapted therefor by spectral analysis, e.g. Fourier analysis or wavelet analysis
• • B—PERFORMING OPERATIONS; TRANSPORTING
  • B23—MACHINE TOOLS; METAL-WORKING NOT OTHERWISE PROVIDED FOR
  • B23K—SOLDERING OR UNSOLDERING; WELDING; CLADDING OR PLATING BY SOLDERING OR WELDING; CUTTING BY APPLYING HEAT LOCALLY, e.g. FLAME CUTTING; WORKING BY LASER BEAM
  • B23K26/00—Working by laser beam, e.g. welding, cutting or boring
  • B23K26/70—Auxiliary operations or equipment
  • B23K26/702—Auxiliary equipment
• • B—PERFORMING OPERATIONS; TRANSPORTING
  • B23—MACHINE TOOLS; METAL-WORKING NOT OTHERWISE PROVIDED FOR
  • B23K—SOLDERING OR UNSOLDERING; WELDING; CLADDING OR PLATING BY SOLDERING OR WELDING; CUTTING BY APPLYING HEAT LOCALLY, e.g. FLAME CUTTING; WORKING BY LASER BEAM
  • B23K31/00—Processes relevant to this subclass, specially adapted for particular articles or purposes, but not covered by only one of the preceding main groups
  • B23K31/12—Processes relevant to this subclass, specially adapted for particular articles or purposes, but not covered by only one of the preceding main groups relating to investigating the properties, e.g. the weldability, of materials
  • B23K31/125—Weld quality monitoring
• • G—PHYSICS
  • G01—MEASURING; TESTING
  • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
  • G01N29/00—Investigating or analysing materials by the use of ultrasonic, sonic or infrasonic waves; Visualisation of the interior of objects by transmitting ultrasonic or sonic waves through the object
  • G01N29/14—Investigating or analysing materials by the use of ultrasonic, sonic or infrasonic waves; Visualisation of the interior of objects by transmitting ultrasonic or sonic waves through the object using acoustic emission techniques
• • G—PHYSICS
  • G01—MEASURING; TESTING
  • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
  • G01N2291/00—Indexing codes associated with group G01N29/00
  • G01N2291/26—Scanned objects
  • G01N2291/267—Welds

Definitions

• 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.
• 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.
• laser beam welding is chosen for high quality joining of materials.
• these systems have to be set-up and calibrated manually with high efforts.
• 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.
• these sensor systems are used as input for the examined cognitive technical system.
• 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.
• acoustic sensors may support existing monitoring methods significantly.
• 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.
• ANN Artificial Neural Networks
• SVM Support Vector Machines
• Fuzzy Logic Fuzzy Logic
• 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.
• 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.
• the acoustic data comprises solid-borne acoustic data or air-borne acoustic data.
• the acoustic frequency features are classified by means of a support vector classifier.
• 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.
• 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).
• Isometric feature mapping Isomap
• PCA Principal Component Analysis
• LLE Locally Linear Embedding
• MDS multidimensional scaling
• only the first component obtained from dimensionality reduction is used to classify the laser process.
• the learned acoustic frequency features are acquired during an initial training, in which a human expert classifies a plurality of acoustic frequency features.
• classifying the acoustic frequency features preferably comprises the detection of a full penetration event through the workpiece.
• 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.
• 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.
• the method further comprises the recording of further sensor data to be fed into the classification process.
• a laser material processing head which comprises a control unit being adapted to perform the method of the present invention.
• the laser material processing head further comprises piezo sensors to be mounted on the workpiece for detecting solid borne-acoustic signals from the workpiece.
• 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. 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.
• 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.
• the techniques used to process the sensor data and the experimental setup are described. Then, two embodiments of the present invention are discussed.
• laser power variation is used to determine changes within the solid-borne acoustics
• the second embodiment aims to use solid-borne acoustics for monitoring purposes.
• 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.
• 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.
• 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.
• the windowed Fourier transform 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 .
• 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.
• 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.
• n 2 1 frequency bands.
• 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.
• C Dimensionality reduction
• Isometric feature mapping Isomap
• PCA Principal Component Analysis
• LLE Locally Linear Embedding
• MDS multidimensional scaling
• 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.
• 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 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.
• 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.
• 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.
• Two different trials were performed to analyze the solid-borne acoustics of laser welding processes.
• 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.
• 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.
• 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.
• 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.
• FIG. 1 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.
• 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.
• PSD power spectral density
• 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.
• 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.
• the Support Vector Machine 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.
• 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.
• 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.
• 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.
• ANN Artificial Neuronal Networks
• 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.
• 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.

Landscapes

• Engineering & Computer Science (AREA)
• Physics & Mathematics (AREA)
• Analytical Chemistry (AREA)
• Biochemistry (AREA)
• Pathology (AREA)
• Signal Processing (AREA)
• Immunology (AREA)
• Health & Medical Sciences (AREA)
• Life Sciences & Earth Sciences (AREA)
• Chemical & Material Sciences (AREA)
• General Physics & Mathematics (AREA)
• Mechanical Engineering (AREA)
• General Health & Medical Sciences (AREA)
• Quality & Reliability (AREA)
• Spectroscopy & Molecular Physics (AREA)
• Mathematical Physics (AREA)
• Optics & Photonics (AREA)
• Plasma & Fusion (AREA)
• Acoustics & Sound (AREA)
• Laser Beam Processing (AREA)

Priority Applications (1)

Applications Claiming Priority (3)

Publications (1)

Family

ID=43446726

Family Applications (1)

Country Status (2)

Families Citing this family (3)

Citations (5)

Family Cites Families (7)

• 2010
  • 2010-09-03 WO PCT/EP2010/005434 patent/WO2011026638A1/en active Application Filing
  • 2010-09-03 EP EP10759806A patent/EP2473314A1/de not_active Withdrawn

Patent Citations (5)

Non-Patent Citations (2)

Also Published As

Similar Documents

Legal Events