Classifications

• • 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
  • B23K26/02—Positioning or observing the workpiece, e.g. with respect to the point of impact; Aligning, aiming or focusing the laser beam
  • B23K26/04—Automatically aligning, aiming or focusing the laser beam, e.g. using the back-scattered light
  • B23K26/046—Automatically focusing the laser beam
• • 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
  • B23K26/02—Positioning or observing the workpiece, e.g. with respect to the point of impact; Aligning, aiming or focusing the laser beam
  • B23K26/03—Observing, e.g. monitoring, the workpiece
• • 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
  • B23K26/02—Positioning or observing the workpiece, e.g. with respect to the point of impact; Aligning, aiming or focusing the laser beam
• • 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
  • B23K26/02—Positioning or observing the workpiece, e.g. with respect to the point of impact; Aligning, aiming or focusing the laser beam
  • B23K26/04—Automatically aligning, aiming or focusing the laser beam, e.g. using the back-scattered light

Definitions

• the invention relates to a method and a device for monitoring a laser processing operation to be performed on a workpiece as well as to a laser processing head with such a device.
• process monitoring systems and sensors are used both in a laser cutting process and in a laser welding process.
• sensors are used to detect the radiation coming from a working or interaction zone determined by the working focus.
• Radiation sensors for observing a plasma forming above the interaction zone and a back-reflection sensor are provided as standard and detect the reflection of the laser from the interaction zone between the laser beam and a workpiece to be machined.
• temperature sensors or infrared sensors are used to monitor the laser processing operation, by which an edge melting and the temperature profile during processing can be monitored.
• the laser processing operation is further monitored by cameras, which may also be sensitive in predetermined wavelength ranges. Due to the image processing of the images taken by the cameras, characteristic values for monitoring the laser processing process, such as parameters with regard to the melted workpiece area, can likewise be obtained.
• the first objective of the monitoring systems is to first classify the processing quality according to the process specification.
• the second goal is to improve the processing quality by controlling and regulating the processes.
• the sensors and cameras used for process monitoring are used by means of the acquired Sensor data and using methods of image processing and data analysis to perform a classification of the current state of the machining process.
• the processes used are individually adjusted to the machining processes.
• the current machining process is classified as insufficient, with appropriate control mechanisms being used to remedy this condition.
• the regulation of the process parameters with respect to the recorded sensor data relates only to the respective measurement data of the corresponding sensors.
• An object of the present invention is to provide a method and apparatus for monitoring a laser processing operation to be performed on a workpiece by which the classification of a laser processing state and thereby the processing quality of a laser processing operation to be performed on a workpiece are improved.
• a method for monitoring, controlling or regulating a laser processing operation to be carried out on a workpiece comprising the following steps: detecting at least two current measured values by means of at least one sensor monitoring the laser processing operation, determining at least two current characteristic values from the at least two current measured values, wherein the at least two current characteristic values jointly represent a current fingerprint in a characteristic space, providing a predetermined point quantity in the characteristic value space, and classifying the laser processing operation by detecting the position of the current fingerprint relative to the predetermined point quantity in the characteristic value space.
• the inventive method additionally comprises a step of Regge at least one process parameter of an associated actuator such that when leaving the current fingerprint from the point set of the characteristic space of at least
• an actuator is activated so that the change of the associated process parameter corresponds to a gradient in the characteristic space, which extends from the fingerprint in the direction of the predetermined set amount in the characteristic space.
• the determination of a current characteristic value from at least one current measured value comprises a method for data reduction or dimension reduction such as a principal component analysis, multidimensional scaling, support vector machines or a support vector classification. Due to the dimensional reduction of the sensor data, it will be possible to
• the classification can be done much faster by a computer, which, for example, a fast control of a laser processing operation can be performed.
• a computer which, for example, a fast control of a laser processing operation can be performed.
• the determination of a current characteristic value from at least one current measured value takes place with the aid of a neural network.
• the measurement data often 25 not directly inferences on a processing situation allow, it is advantageous if the predetermined amount of points within the characteristic space is determined by means of a learning process.
• the gradient field of the parameter space is determined as a function of the process parameters in different regions at the points in the characteristic space which are representative of the gradient with respect to the gradient, the gradient of the characteristic space depending on a process parameter is determined by a variation of the process parameter at a predetermined point of the characteristic value space.
• the at least one sensor is selected from a group which comprises at least one photodiode with filters for specific wavelengths, body and airborne sound pickups, and at least one camera unit with a corresponding surface illumination includes.
• the camera images acquired by the at least one camera unit are recorded with different exposure times and to be charged to one another via the high-dynamic-range method.
• the at least one actuator is selected from a group which controls the laser power, a speed control of the machining head relative to the workpiece, a control of the focal position of the machining laser beam, a control of the distance of the machining head Workpiece, and a control of the lateral offset comprises.
• an apparatus for carrying out the method according to the invention which has at least one sensor for monitoring the laser processing operation, which is suitable for detecting at least two current measured values, a data processing unit for determining at least two characteristic values from the at least two current measured values for creating a current one Fingerprint in a characteristic space, a storage unit for storing a predetermined amount of points within the characteristic space, and a classification unit, which is suitable for evaluating the laser processing operation by detecting the position of the current fingerprint relative to the predetermined set amount in the characteristic space.
• the device further comprises a control unit for controlling at least one process parameter of an associated actuator such that when leaving the current fingerprint from the point set of the characteristic space the at least one Actuator is activated so that the change of the associated process parameter corresponds to a gradient in the characteristic space, which extends from the fingerprint in the direction of the predetermined point amount.
• the at least one sensor is selected from a group which comprises at least one photodiode with filters for specific wavelengths, body and airborne sound pickups, and at least one camera unit with a corresponding surface illumination.
• the at least one actuator is advantageously selected from a group which comprises a control of the laser power, a speed control of the machining head relative to the workpiece, a control of the focus position of the machining laser beam, a control of the distance of the machining head to the workpiece, and a control of the lateral offset ,
• a laser processing head for processing a workpiece by means of a laser beam is provided according to the invention, which comprises the device according to the invention.
• FIG. 1 shows a flowchart with the essential components of the machining process of a workpiece according to the method according to the invention
• FIG. 2 shows an overview of the sensors used in the method according to the invention for monitoring and detecting the laser processing process
• FIG. 3 shows a greatly simplified schematic view of the components used in a machining process according to the invention
• FIG. 4A shows a greatly simplified schematic view of part of the actuators used in the method according to the invention in a laser beam welding process
• FIG. 4B shows a greatly simplified schematic view of part of the actuators used in the method according to the invention in a laser-cutting process
• FIG. 5A shows a flowchart of the generation of a fingerprint according to the method according to the invention using linear as well as non-linear dimension reducers
• FIG. 5B shows a flow chart of the generation of a fingerprint according to the method according to the invention using a neural network
• FIG. 6A is a flowchart of the classification process according to the method of the invention using linear and nonlinear dimensional reducers.
• FIG. 6B shows a flow chart of the classification process according to the method according to the invention using a neural network
• FIG. 7 is a schematic diagram illustrating an error detection method
• FIG. 8 shows a flow diagram which illustrates the learning of the fingerprints or the characteristic values or features according to the invention
• FIG. 9 is a flow chart of the dimensional reduction method according to the invention.
• FIG. 10 shows a flow chart of the assessment of the current machining process according to the invention
• FIG. 11 is a flow chart of the estimation of new control parameters according to the invention.
• FIG. 12 is a schematic view of a camera image processed with an HDR method according to the invention
• FIG. 13 shows a block diagram of an HDR image sequence processing according to the invention.
• FIG. 14 shows a flow chart of a classification process using a reinforcement learning method in a laser processing operation according to the invention
• FIG. 15 is a flow chart of a classification process using a discriminant analysis method in a laser processing operation according to the invention.
• FIG. 16 is a flowchart of a control operation by means of set values obtained by dimensional reduction in a laser processing operation according to the invention.
• a cognitive laser material processing system which has cognitive abilities through the use of machine learning and self-learning algorithms.
• the associated method according to the invention can be used in laser material processing for process observation, process control, and process control.
• a system can have two types of cognitive abilities: First, it appears to an external observer as if the observed system had cognitive abilities, such as the ability to learn and improve independently. Second, the system realizes cognitive abilities in a similar way to a natural organism, such as the human brain.
• the system according to the invention possesses cognitive abilities such as learning as well as the independent recognition and improvement of errors that are used in laser material processing.
• the use of cognitive skills is particularly advantageous in the field of laser material processing. Machining processes such as cutting or joining workpieces are very different from process to process. So far, it is known to set each process manually individually first. After setting the process parameters, the process is only observed and adjusted accordingly manually. For example, if a next batch of workpieces is dirty or the workpiece thickness deviates from the previous batch of workpieces, the process often has to be readjusted manually. An automatic adaptation to process changes was either not at all possible or only possible to a very limited extent. In fact, the requirements of vehicle manufacturers who want to produce several vehicles in a production line are such that the production systems can adapt quickly and adaptively to the machining processes.
• the fast learning of machining processes and the recognition, improvement and avoidance of errors during the machining are requirements that are met by the cognitive capabilities of the machining system according to the invention.
• Fig. 1 is a flowchart of the method according to the invention with its essential components is shown schematically, which are explained below step by step.
• all relevant information of the machining process is detected according to the invention with a sensor system having at least one sensor.
• the sensors used are used to obtain a multiplicity of measured values and information about the process in order to be able to determine features, process images, process characteristics or clear fingerprints of the process, which are referred to below as characteristic values, from the measurement data of the sensors monitoring the processing process.
• the determination is carried out in particular by calculating or another suitable, preferably electronic processing of the measured values.
• FIG. 2 An overview of the sensors used according to the invention is shown in FIG. 2 and a structure of a laser processing system according to the invention with the corresponding sensors is shown in FIG.
• sensors for detecting used by body and airborne sound In addition to already known sensors for monitoring a laser processing process, according to the invention additionally sensors for detecting used by body and airborne sound.
• sensors for body and airborne sound For sound recording, it is expedient to use at least two sensors each for body and airborne sound.
• the sensor signals for body and airborne sound are filtered, amplified and correspondingly scanned depending on the process in preprocessing. For airborne sound recording different directional characteristics are suitable.
• the locations of the sound sources and the direction of propagation can be calculated. This can also reduce noise from non-relevant sources and background noise, or apply methods such as Active Noise Cancellation.
• Sensors for emission detection of specific wavelengths which are preferably photodiodes that are sensitive for a specific wavelength range are also provided in the laser processing head.
• optical bandpass filters for selecting specific wavelength ranges may additionally be arranged in front of the corresponding photodiodes. The measured values of these sensors are also recorded and sampled.
• cameras which observe the laser processing operation and in particular the laser processing zone are used for data acquisition.
• an in-process camera can be used, whose observation beam path is coaxially coupled into the beam path of the working laser in the processing head, so as to image the laser processing zone.
• a camera can also record the processing process outside the processing head.
• a leading camera, called a pre-process camera, and a trailing camera, called a post-process camera, can also capture the laser processing process.
• Various workpiece lighting concepts are suitable for camera acquisition, depending on the machining process.
• illumination light-emitting diodes which are inexpensive and can radiate in a wide wavelength range
• lasers are used in different wavelengths with appropriate optics for focusing on the camera section on the workpiece surface.
• data processing methods such as "region of interest”, “qualas”, or geometry data evaluation are particularly suitable and preferred.
• a high dynamic range (HDR) method is used, which advantageously increases the contrast ratio of the captured camera images.
• the inventive method is not limited to the use of the plurality of sensors, but already using only one sensor, such as the in-process camera, can be performed.
• control program In laser material processing, a control program is normally designed manually for all involved actuators. During the process, this control program is only monitored via process monitoring or adjusted with firmly defined control circuits such as capacitive distance sensors during laser cutting.
• the laser beam power, the distance between the machining head and the workpiece, the speed of the machining head relative to the workpiece, and the position of the focal point of the machining laser beam are controlled.
• laser cutting in the processing method additionally controls or regulates the supply of process gas.
• control signals can be modulated in their intensity with a certain frequency, for example, a modulation of the laser radiation intensity between 90 and 100 percent. Since the control signal is known, the system response via the sensor data can provide insights into the process for example, how to recover a gradient field of the characteristic value space as a function of the process parameters in different measuring ranges.
• the controls can be realized via corresponding linear axes, robot control or other control interfaces.
• the method according to the invention is not limited to the use of the multiplicity of actuators, but can already be carried out using only one actuator, for example a laser power controller for laser welding or a process gas controller for laser cutting.
• the technical cognition has to be abstracted from the sensor data, so that the system according to the invention can autonomously make decisions for the actuation of the actuators.
• the system can be trained by an operator of the system and is self-learning.
• the invention provides that the system independently already knows or acquires and learns the essential characteristic values from all the sensors used and then makes decisions for the process control.
• three stages of the method according to the invention will be illustrated, namely the learning of the process environment, the classification of the current process result and the control or regulation of the process.
• homing or test processing are necessary. Each machining process has a desired result and a different one.
• the test processing or homing must contain both results and ideally also the transitions, as well as the reaction on of the system on the process control. If, for example, a weld is to be achieved in the lap joint of stainless steel with a defined weld width of X mm and a length of Y cm, then at least one homing run must be carried out in which at least one process parameter is varied so that both the defined and the homing are determined the definition violation is contained in both directions of the process parameter.
• the human system operator can carry out a reference run with an increasing laser power as process parameter, in which process the upper and lower definition limit occurs and is exceeded.
• a reference run can start with a laser power which does not yet cause any welding.
• the laser power is controlled to increase continuously until suturing occurs. This process is monitored and used to learn the process environment using the described process sensors, which record the corresponding measurements.
• Another example concerns production problems between two batches of greasy and non-greasy workpieces.
• the definition limits must be included for learning during the reference run.
• the operator tells the cognitive laser material processing system where the definition limits lie, so that the system according to the invention can learn to distinguish between the areas.
• linear and non-linear dimensional reducers and manifold learning methods such as Principal Component Analysis (PCA), MDS (Multidimensional Scaling), LLE (Locally Linear Embedding), and SVM (Support Vector Machines) can be used to understand the process environment. These methods can be used both in combination and alone. To learn the process environment can continue a discriminant analysis are used, as described below.
• PCA Principal Component Analysis
• MDS Multidimensional Scaling
• LLE Long-Linear Embedding
• SVM Small Vector Machines
• KNN artificial neural network
• the procedure is different, since the network is trained here and the learned information is then available in the network, which can then classify the result.
• the initial neurons thus initially provide a classification based on the trained data. Based on this classification can then be regulated.
• the current process result must be detected, compared with the previously learned target range, which can be regarded as a point quantity in the parameter space, and, if necessary, the process parameters are adapted, as shown in FIGS. 6A and 6B.
• the process parameter adaptation can and should already take place before exiting the target range.
• the predetermined point quantity which is used for the regulation of the system can be adapted such that in a control case the current fingerprint of the sensor system already leaves the predetermined point set at a time when the fingerprint enters an edge region of the desired range.
• the cognitive laser material processing system has already stored in the database of a memory the learned process environment, the learned features or fingerprints in the form of a vector or a matrix.
• the process currently used by the process NEN measured values of the sensors must first be reduced in the amount of data and brought for comparison in the same data space, so the characteristic space, such as the feature vectors or fingerprints, thus a current fingerprint is obtained as a reduced sensor data vector or matrix in the characteristic space, with the learned point set is compared in the characteristic value space.
• the probability can be obtained that the currently detected data point is closest to a certain feature point. It is known here whether this feature point is still within the desired range, furthermore the probably necessary correction of the process parameter is known.
• the classification of the current process result by means of neural networks is carried out by the trained network.
• the classification result is whether the process is still within the target range and with which trend the process parameter is to be adapted.
• the control of the process according to the inventive method is carried out in the following manner.
• the control unit already knows the direction and the strength with which the corresponding actuators must be activated.
• Various control methods can be used. For example, the minimization of the geodesic distance between desired feature vector and result vector or a control method with Kalman filter and minimization of the mean square error can be used.
• the tendency to regulate from the multidimensional feature spaces or characteristic dreams can be determined via the "Support Vector" classification. The controller must not exceed the previously defined safety range.
• the invention can be applied in several process variants, some of which are presented here.
• the cognitive laser material processing system uses the PCA, Principal Components Analysis or a combination of the other methods to calculate the proposed dimension reducers from the sensor data.
• the operator now tells the system where to place the workpiece - no garden has arisen.
• the cognitive system can then calculate the corresponding component, a vector or matrix, from the information about where the ridge has been created at the cutting edge and at which points corresponding main components were calculated from the sensor data, the collected features or the fingerprint for contains the formation of burrs. From the current sensor data in the further operation of the system can then be calculated with matrix vector algebra during the process and the operator to see whether the learned error has occurred.
• the same procedure can be used to detect, for example, effects in laser beam welding or laser cutting: wrong friend, suture, penetration, X cut, average state, cut edge roughness, combustion effects, weld width Y, weld in status, weld through status, bond tread status, gap in lap, Gap in butt joint, lateral offset, ejects, pores, holes.
• the invention can also be used to simplify a batch change which previously required an adaptation of the laser material processing system.
• the workpieces of the new batch have slightly changed properties, eg material thickness or degree of contamination.
• It is again carried out first a learning phase and then a classification phase. After the classification phase, a control process can already be realized.
• new control parameters for a process change that occur, for example, as a result of a batch change.
• the learning phase according to FIG. 8, the measured values of the process sensor system are detected by a reference travel.
• constant process control parameters are again set, except for a control parameter that is varied.
• the laser power can be increased steadily during reference travel.
• the recorded data are processed by the cognitive laser material processing system with dimension reducers, cf. FIG. 9.
• the output data of each sensor used are first filtered with a corresponding low-pass filter. Then the n principal components are output via the Principal Component Analysis. The data are then normalized and freed from the mean.
• the corresponding features or fingerprints and their mapping rule are stored in a database for feature ablation provisions.
• the operator of the system now defines an area on the workpiece that corresponds to the desired result. This definition is transformed into a vector with which a classifier can be trained. In order to be able to perform a classification, Support Vector Machines is used in this process.
• a Support Vector Classification method is used. It describes a mathematical method for distinguishing desired and undesired process results, which is performed by a multi-dimensional separation of the feature space, based on the specifications of the operator.
• the database of feature mapping rules describes the mapping rule, and the classification database describes the separation of the feature spaces.
• the cognitive laser material processing system monitors the machining process according to the previously learned operator wishes.
• the sensor data is dimensionally reduced based on the specifications of the particular feature mapping rules.
• the output data are located in the predetermined feature space or feature value space.
• the classification data learned by the operator through the Support Vector Classification procedure is used to assess the current machining process. It can be judged whether the current process is within the user-defined desired range and which tendency is to be taken over a probability for the process control parameter to control the process.
• the estimation of new control parameters or process parameters for small process changes by a batch change will be described. If the machining process is changed for a certain duration, e.g. by slightly changing the workpiece characteristics during a batch change, the new control parameters can be estimated. For this purpose, in addition to the previous homing 1, a new homing 2 must be performed. Homing 1 and 2 used the same control parameters.
• the sensor data or the measured values of the sensors of reference travel 2 are again reduced in size.
• the mapping rules are now applied to the recorded sensor data of Homing 1.
• the occurrence probabilities of the characteristics from reference travel 1 during reference travel 2 are calculated.
• the cognitive laser material processing system can thus calculate, from the position on the workpiece or from the control parameters used here and the occurrence probabilities of the features, which control parameters in the new process will produce a result very similar to or almost the same as in the previous machining process.
• features are obtained from the process data as in the previously described methods. These features are classified by initial and periodic homing by the operator, with appropriate assessment of whether the control parameter should be adjusted.
• the corresponding characteristics and the associated classifications are stored in a database, possibly with an adjustment proposal.
• the operator therefore assesses the system at regular intervals and thus trains it.
• the system can thus first determine whether the current process result is still in the specified feature space and whether the system should carry out an adaptation of the control parameters.
• the learned features and customization suggestions thus become more and more over time and the system becomes better and better in processing. Similar It is possible to recalculate characteristics and adaptation proposals in order to avoid a flood of features.
• HDR High Dynamic Range
• an imaging sensor is either scanned several times, ie at least twice, at different times per image or multiple images, ie two, three or more images with different exposure times or with multiple cameras and then computed with each other to at least one image .
• This procedure enables a picture, image sequence or video recording, which at the same time makes the surrounding processing area, the process lighting and the vapor capillary or keyhole visible in an image.
• the ranges mentioned are distributed in the intensity values in an image acquisition of laser processing processes in a wide range, which can be made visible in an image by said method.
• a picture or image sequence thus created is displayed adapted via a gray value or tone mapping method.
• a plurality of images or pixel arrays are offset with one another.
• the different images can be created by multiple scanning of an imaging sensor or by simultaneous image acquisition with multiple cameras or by sequential image acquisition with a camera, but different exposure times, called multi-exposure technique.
• the billing of individual images can be done on different types of procedures. In the simplest case, this includes adding up and averaging the individual image values of several images of a sequence of images from at least two image recordings. For better image acquisition, the image values or pixels from an image sequence can be averaged out of at least two image recordings. Either an entropy method can be used as the weighting method, for the weighting according to the information content, or a weighted averaging can be carried out taking into account the camera response function. For this, a conclusion must be drawn on the real or realistic radiation energy per area, which is given by the following function:
• the weighting for the individual radiation energies is then:
• i is the image index of an image sequence of several image recordings
• j the pixel position
• tj the exposure time or scan time of the image capture iyy the intensity value of the pixel of the image capture i at the position j
• T 1 O the inverse Camera Response Function
• X j the estimated radiation energy per area at pixel position j
• Wy the weighting function of the reliability model.
• the invention explicitly relates to the use of these illustrated HDR image processing methods in processing methods such as cutting or joining of materials, in particular with laser processing heads and / or the process monitoring system according to the invention connected thereto.
• any sensor that enables a sensor data output can be used as the sensor system.
• these are, for example, microphones or structure-borne sound pickups, cameras, photodiodes, buttons, technical evaluation and monitoring signals and Aktorikparameter, such as the laser power.
• PCA Principal Component Analysis
• ICA Independent Component Analysis
• Wavelet Analysis Fourier
• Fast Fourier Fast Fourier
• Laplace Analysis feature and object recognition algorithms. driving, Locally-Linear Embedding, Artificial Neural Networks, Multidimensional Scaling and much more.
• the reduced amount of data can be interpreted as a point cloud of a multi-dimensional space obtained from a higher-dimensional space.
• By reducing the data it is possible to compare them in finite time with previously recorded and classified or learned data sets. In this classification, it can be determined whether the new sensor data are similar to already recorded sensor data and this similarity is assigned a probability. If a defined threshold for a likelihood of similarity of a previously recorded amount of data is exceeded, then the solution or control or regulation approach previously stored thereon can be tracked. If the threshold for a likelihood of similarity to previously learned amounts of data is exceeded, the system has a new situation.
• the behavior for a new situation can either be learned by inquiring from a human operator or can be tried out of the previous data and solution strategies according to the similarity principle.
• self-learning algorithms are used, which then check after a target, after trying out a self-developed approach, whether a goal has been achieved and evaluate the chosen approach accordingly.
• Support Vector Machines Support Vector Classification, Fuzzy Logic, Information Fuzzy Networks, Fuzzy K-Nearest Neighbor Classifier, K-Nearest Neighbor Classifier, Reinforcement Lear - ning, Bayesian Networks and Bayesian Knowledge Databases, Naive Bayes Classifiers, Hidden Markov Chains, Artificial Neural Networks and Backpropagation, Regression Analysis, Genetic Programming or Decision Trees.
• the solution strategy resulting from the classification, or a controller or actuator control can be carried out simply, but it can also control the type of data acquisition. If, for example, no threshold is reached for a known amount of data, then the type of data acquisition can be changed. For example, this can be done by adapting a wavelet analysis. to new frequency ranges or by switching from PCA to ICA.
• High Dynamic Range method (HDR method)
• An HDR method may be used to calculate a higher contrast ratio from multiple captured images or image matrices and vectors having different contrast ratios. For this purpose, when taking a picture or observing a scene, several pictures with different exposure times can be taken, from which subsequently a picture or a picture sequence with improved contrast ratio can be calculated. In order to produce a sequence of images with different contrast ratios, several images with different exposure time can be recorded, according to the so-called multi-exposure method.
• the pixel values can also be scanned repeatedly during an exposure time. In this way, an image sequence with different contrast ratios during an exposure time is created.
• an imaging sensor can also be scanned repeatedly during an exposure time.
• Charges representing the pixels are retrieved once and thereafter can not be retrieved a second time.
• Techniques, such as non-destructive reading include non-distructive readout
• NDRO Multi Slope or Single Slope Readout or Cooled Imager or Charge Injection Imaging (CIS) or Thin-Film on CMOS (TFC) or Active Pixel Sensor (APS) or Single-slope or Correlated Double Sampling (CDS) that allows multiple interrogation of a charge, for example, on a CMOS chip, during a single exposure period without the interrogation of the interrogated charge value.
• these techniques can be used for observing a laser processing process in order to realize an observation or control method, whereby it is possible by the HDR method to carry out simultaneously the process emissions, the vapor capillary, the weld pool, the weld seam geometry during a laser welding operation to be carried out.
• RL Strengthening Learning or Reinforcement Learning
• RL refers to a field of machine learning. It describes procedures in which systems or agents, English, action apply to environments to maximize a reward. RL finds mapping rules or procedures (Engl, policy) for one or more system states or states on system action plans or actions. The methods of RL can be used according to the invention for self-improving control and observation of laser processing processes.
• FIG. 14 shows a possible procedure for how RL can be integrated in a laser processing process.
• the values to be learned are symbolized by the matrix Q.
• the Q matrix consists of the components QS1, QSn, QSA, QDR, QR1, QRm, these can contain one or more values. These components are initialized with a start value and optimized according to an RL procedure. This optimization takes place by performing an action, evaluating it with a reward function, and modifying the values of the Q Matrix. Comparable to a theater where an actor is judged by a critic, and the actor adjusts his actions. As described above, in a reference run or by a learning phase, a point cloud with appropriate classification can be obtained by a human expert.
• the characteristics or point clouds or characteristics or fingerprints or sensor measured values which represent the desired process result are therefore stored in this.
• This can be realized by a support vector machine or another classification method.
• This can be a reward function that the RL method works for.
• the Q matrix is thus optimized after this man-taught reward function.
• weighting values or setting parameters can be learned and optimized, such as the weighting different sensors among each other (QSl, QSn), the selection of special features that are used for control or observation (QDA), the selection of setpoints for various control methods (QDR), or the controller setting parameters such as proportional, P-share, integrated , I share, and differentiated, D share (QRl, QRm).
• MDP Markov Decision Process
• AHC Q Learning Adaptive Heuristic Critic
• SARSA State-Action-Reward-State-Action
• ART Adaptive Resonance Theory
• MCA Multivariate Analysis
• EM Expectation-Maximization Algorithm
• Radar Dial Basis Function Network Time Series Prediction
• ATR Automatic Target Recognition
• RBF Radial Basis Function
• a discriminant analysis (DA) or linear discriminant analysis or linear discriminant analysis (LDA) as well as Fisher's linear discriminant is a statistical analysis method which has a similar principle to the principal component analysis already described. In contrast to the principal component analysis, the DA also considers the class affiliation of a classification. Also, DA can alternatively be used for dimensional reduction in the method according to the invention, but at the same time represents a combination of dimensional reduction and classification method.
• sensor data can be acquired, reduced in dimension and classified with previously learned data using a method as already described.
• the classification result can then be used as the basis for an actual value calculation for one or more controllers with learned setpoint values for the control of one or more actuators or control parameters.
• the DA can be combined with other material reduction methods in laser material processing, for example, a main component can be used first Be carried out analysis and then carried out a DA.
• This also applies to the other dimension reduction methods already described, which reduce a sensor data input vector of dimension Y to a dimension X with X ⁇ Y.
• the combinations may differ for the respective sensors.
• the already mentioned Independent Component Analysis which extracts features for statistical independence, is particularly suitable for acoustic sensors and the principal component analysis for imaging sensors.
• Further dimension reduction methods can be used according to the invention in a described laser material processing system: kernel principle component analysis (Kernel Principle Component Analysis), locally linear embedding (LLE), Hessian LLE, Laplace proper name and map (Laplacian own maps) , Local Tangent Space Alignment (LTSA), Semi-Definite Embedding (SDE), Maximum Variance Unfolding (MVU), Curvilinear Component Analysis (CCA), Data-driven High-dimensional Scaling (DD-HDS), Autoencoders, as a special variant of a feed -Forward Artificial Neural Network, Boltzmann Machines as well as all procedures similar principle.
• a principal component analysis or other dimensional reduction methods or a feature extraction or an HDR method can also be carried out on a cellular neural network (CNN) integrated in an image acquisition unit in a laser processing system for particularly fast data processing.
• CNN is a parallel calculation method similar to an artificial neural network.
• a laser processing process can also be controlled directly with set values from a dimension reduction for faster data processing, a classification can then be used to determine the best set values with an optimization of a signal to signal to noise ratio. In this way, very high control cycles can be realized, while high adaptability by taking into account the learned classification results.

Landscapes

• Physics & Mathematics (AREA)
• Optics & Photonics (AREA)
• Engineering & Computer Science (AREA)
• Plasma & Fusion (AREA)
• Mechanical Engineering (AREA)
• Laser Beam Processing (AREA)
• Length Measuring Devices By Optical Means (AREA)

Priority Applications (11)

Applications Claiming Priority (8)

Publications (1)

Family

ID=41666804

Family Applications (2)

Family Applications After (1)

Country Status (11)

Cited By (12)

Families Citing this family (85)

Citations (5)

Family Cites Families (23)

• 2009
  • 2009-11-20 CN CN200980147217.2A patent/CN102281984B/zh active Active
  • 2009-11-20 WO PCT/EP2009/008293 patent/WO2010057661A1/de active Application Filing
  • 2009-11-20 CA CA2743522A patent/CA2743522C/en not_active Expired - Fee Related
  • 2009-11-20 US US13/130,426 patent/US9492886B2/en active Active
  • 2009-11-20 EP EP09760109.0A patent/EP2365889B1/de active Active
  • 2009-11-20 CN CN200980155381.8A patent/CN102292187B/zh active Active
  • 2009-11-20 RU RU2011125317/02A patent/RU2529135C2/ru active
  • 2009-11-20 KR KR1020117014139A patent/KR101700896B1/ko active IP Right Grant
  • 2009-11-20 RU RU2011125341/02A patent/RU2529136C2/ru active
  • 2009-11-20 US US13/130,499 patent/US9056368B2/en active Active
  • 2009-11-20 KR KR1020117014138A patent/KR101787510B1/ko active IP Right Grant
  • 2009-11-20 MX MX2011005335A patent/MX2011005335A/es active IP Right Grant
  • 2009-11-20 JP JP2011536789A patent/JP5763542B2/ja active Active
  • 2009-11-20 CA CA2743518A patent/CA2743518C/en not_active Expired - Fee Related
  • 2009-11-20 WO PCT/EP2009/008294 patent/WO2010057662A1/de active Application Filing
  • 2009-11-20 EP EP09760110.8A patent/EP2365890B1/de active Active
  • 2009-11-20 JP JP2011536788A patent/JP5762968B2/ja active Active
  • 2009-11-20 MX MX2011005336A patent/MX2011005336A/es active IP Right Grant
• 2011
  • 2011-06-20 ZA ZA2011/04527A patent/ZA201104527B/en unknown
  • 2011-06-20 ZA ZA2011/04525A patent/ZA201104525B/en unknown
• 2012
  • 2012-06-14 HK HK12105813.0A patent/HK1164786A1/zh not_active IP Right Cessation

Patent Citations (5)

Non-Patent Citations (1)

Also Published As

Similar Documents

Legal Events