WO2011083087A1

WO2011083087A1 - Method for processing workpieces by means of a cognitive processing head and a cognitive processing head using the same

Info

Publication number: WO2011083087A1
Application number: PCT/EP2011/000044
Authority: WO
Inventors: Stork Genannt Wersborg Ingo
Original assignee: Precitec Kg; Precitec Itm Gmbh
Priority date: 2010-01-08
Filing date: 2011-01-07
Publication date: 2011-07-14
Also published as: DE202011110730U1; DE112011100192T5

Abstract

The present invention relates to a method for processing workpieces by means of a cognitive processing head, comprising the steps of performing an initial processing process of the initial workpiece type by acting on learned knowledge; and making an autonomous decision in a previously unknown situation for adapting the learned knowledge to a secondary processing process of a secondary workpiece type.

Description

Method for processing workpieces by means of a cognitive processing head and a cognitive processing head using the same

Description

The present invention relates to a method for processing workpieces by means of a cogni- tive processing head and a cognitive processing head using the same. In particular, the present invention is directed to transferring inspirations from natural cognition into realtime process control for industrial production systems.

Industrial production uses a high number of robots and complex technical systems to build different products. Furthermore, manufacturing has to satisfy two needs at once - to further increase mass production and to create more flexibility in individualized versions of complex products, referred to as mass customization. This need may be partly solved by highly adaptive production systems, which can learn how to fulfil manufacturing tasks or how to monitor and control them using machine learning. Such adaptive and autonomous production systems would save costs and labour by enabling easier changes of assembly lines, improving product processing quality and preserving environmental resources. Laser material processing (and laser beam welding in particular) is an example of a complex and hard-to-control production process frequently used, for example, in car or airplane assembly lines. However, to control production processes in real-time, the execution times of many machine learning algorithms are too long. It is an object of the present invention to provide a method for processing workpieces by means of a cognitive processing head and a cognitive processing head using the same, which provide an architecture or design of technical cognitive methods suitable for controlling at least one parameter in industrial production processing tasks utilizing sensors and actuators in a reasonable timeframe. This object is solved by a method for processing workpieces by means of a cognitive processing head according to claim 1 and by a cognitive processing head according to claim 15. Further advantages, refinements and embodiments of the invention are described in the respective sub-claims. In many cases, if a production system is installed or reconfigured, qualified professionals are present. Herein, the technical cognitive architecture according to the present invention incorporates the possibility to first learn from a human expert and then to learn without supervision or to make decisions on its own. Furthermore, the architecture of the present invention could be applied to problems of laser material processing or robotic welding with laser beams. Laser beam welding is a suitable showcase for production processes which require a high level of precision, using many individually weak sensors for data inputs, and requiring quick adaptation to complex system configurations and real-time process control. The present invention aims to offer a data analysis concept that will satisfy the need for different production scenarios, in particular in view of adaptive and autonomous laser material processing. Many other architectures and frameworks towards cognitive technical systems and machine learning approaches capable of controlling production systems have been demonstrated.

To fulfil the above object, the present invention provides a method for processing work- pieces by means of a cognitive machining or processing head, comprising the steps of performing an initial machining or processing process on a workpiece of an initial workpiece type by acting on learned knowledge; and making an autonomous decision in a previously unknown processing situation for adapting the learned knowledge to a secondary machining or processing process on a workpiece of a secondary workpiece type. In a further embodiment, the present invention provides a method for processing work- pieces by means of a cognitive machining or processing head, cognitive laser machining head, or cognitive laser processing head, comprising the steps of abstracting relevant sensor information concerning or related to an initial machining or processing process on a workpiece of an initial workpiece type; learning from a human expert how to control the initial machining or processing process on a workpiece of the initial workpiece type; performing an initial machining or processing process on a workpiece of the initial workpiece type by acting on learned knowledge; and making an autonomous decision in a previously unknown processing situation for adapting the learned knowledge to a secondary machining or processing process on a work- piece of a secondary workpiece type.

Preferably, the processing of the workpieces comprises a laser welding or cutting process. According to a preferred embodiment of the present invention, the abstracting of relevant sensor information may preferably comprise an initial test run, in which an actuator parameter is altered and the sensor information is processed by a feature extraction technique to embed the sensor data input to a learned feature space.

Preferably, the relevant sensor information comprises information of sensors includ- ing a high-speed camera, sensors for solid-borne and air-borne acoustics, and three photodiodes recording different process emissions in different wavelengths.

According to an advantageous embodiment of the present invention, the feature extraction technique comprises a linear or non-linear dimensionality reduction processing of the sensor data input. It is preferred that the learning from a human expert how to control an initial machining or processing process comprises controlling of actuators including a linear drive or robotic positioning actuator or an actuator for setting a laser power.

According to an embodiment of the present invention, learning from a human expert comprises classification techniques such as Support Vector Machines, Artifi- cial Neural Networks, and Fuzzy k-Nearest Neighbour.

Preferably, performing an initial machining process comprises a combination of a PID-control with classification results.

In a further embodiment of the present invention, performing an initial machining or processing process comprises performing laser power controlled overlap welds, in which too large gaps between two sheets to be joint are monitored. In an advantageous embodiment of the present invention, making an autonomous decision includes a novelty check on the basis of the trained data.

Preferably, if the novelty check is positive, a further test action is performed to classify the secondary workpiece type with the previously trained features by means of an supervised or unsupervised learning procedure.

In an embodiment of the present invention, the secondary workpiece type is preferably related to a different material thickness of the workpiece.

Preferably, the supervised learning procedure includes reinforcement learning.

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

Herein, the machining or processing head of the present invention preferably comprises a laser optic for laser processing of workpieces, actuators including linear drive or robotic positioning actuators, an actuator for setting a laser power, and sensors including a high-speed camera, sensors for solid-borne and air-borne acoustics, and three photodiodes recording different process emissions in different wavelengths.

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 illustrates an architecture for real-time adaptive process control of actuators in industrial production systems according to the present invention;

FIG. 2 shows an architecture design for real-time adaptive process control of laser welding according to the present invention; FIG. 3 A shows an In-process image taken with a co-axial camera;

FIG. 3B shows an image of a laser welding process with acoustic sensors; FIG. 4 shows a diagram of features and probability related to a gap between two sheets of stainless steel;

FIG. 5 shows a diagram related to a probability of a gap between two sheets of stainless steel for a workpiece with a number 151, wherein the real gap is indicated; and

FIG. 6 shows a diagram related to a mapping of features from a reference workpiece onto workpieces of different thicknesses, wherein the shaded zone indicates the usual laser welding process operation area. First of all, the inspiration from natural cognition is identified in a first section. Then, an architecture for real-time process control according to the present invention is introduced. This concept is transferred and experimentally validated for laser beam welding in the last section.

Natural cognition in its complex variations over species and individuals is not yet fully understood. However, recent results may be an inspiration for the design of cognitive technical systems.

Memory coding units have been identified in the hippocampus of mice brains. Events of cognitive importance are not memorized in every detail but in a selective representation. The brain seems to abstract information by decoding it with activation patterns. Sensations then create those activation patterns enabling the mouse to identify previously learned situations. Furthermore, these characteristics not only appear to exist and be alike between individual mice, but also between different species. It has been shown in human brain activity that patterns of neural activation are associated with thinking in different semantic categories. A functional magnetic resonance imaging combination with a computational model, incorporating dimensionality reduction and classification techniques, is described that can display if a test person thinks of previously memorized objects, such as tools. Once the neural activation pattern of a test person is learned, this method can visualize if a different test person is thinking of a hammer or a screw driver. Groups of nerve cells in different combinations appear to react to certain events or thoughts encoded by specific patterns. It seems as if natural cognition is organized with abstracted information or the mentioned patterns representing real-world events in combination with a categorization. Information from the real world is reduced to signals with patterns enabling the brain to distinguish and recognize events. These cognitive capabilities empower humans and other natural cognitive organisms not only to learn a lot but also to react fast. This is also highly desirable in technical systems.

If natural cognitive capabilities can be modelled with mathematical data reduction techniques combined with classifiers, it should also be possible to realize similar capabilities by computation. If neuroscientists can move a cursor on a screen just by thought, the computational methods may also be useful to improve complex controlling tasks in industrial production.

In the following, the technical cognitive architecture for real-time process control of the present invention should be presented. A natural cognitive capability is to abstract relevant information from a greater magnitude and to differentiate categories within this information. Transferring this concept from natural cognition to the world of mathematical data analysis, a combination of data reduction techniques and classification methods can possibly be used to achieve something with similar behaviour. In industrial production, many manufacturing processes can be con- sidered as black box model, focusing on the in and outs of the box rather than what happens inside. The connections to the black box for production systems often are sensors and actuators. Sensors such as cameras, microphones, tactile sensors, and many more are provided to monitor production processes. Furthermore, the systems need actuators such as linear drives or robotic positioning to interact with the real world. For every production process these actuators have to be parameterized.

To learn how to adaptively control at least one parameter of such production systems, many combinations of self-learning algorithms, classifying techniques, knowledge reposi- tones, feature extraction, dimensionality reduction and manifold learning techniques have been examined. Further different controlling techniques, open and closed-loop, with multiple different sensors and actuators have been investigated. After many simulations and experiments, a simple architecture according to the present invention demonstrating how to combine these techniques has been finally found, which has been proved to be successful and reliable at least for laser welding. However, the laser welding process may be interpreted as black box, which may possibly be applicable to other kinds of production processes as well.

In figure 1 , a diagram illustrates an architecture or method of the present invention being suitable for adaptive process control. The diagram describes the unit communication and information processing steps. The architecture or method of the present invention may be designed for four modes of usage or method steps: first, abstracting relevant information; second, receiving feedback from a human expert how to control processes with other words supervised learning; third acting on learned knowledge; and fourth, controlling autonomously processes in previously unknown situations. First of all, the step of abstracting relevant information is discussed in detail. In natural human cognition, we abstract or absorb information from all that we hear, feel, and see. So in general we only remember the most interesting things. Inspired by this, a technical cognitive system may also abstract relevant information from a production process. Working with abstracted features instead of raw sensor data has certain advantages such as the fol- lowing: many weak sensor signals may be transferred to fewer and more reliable feature. Additionally, in order to realize real time process control it is necessary to reduce the incoming sensor data in volume, because greater data amounts may have significant influence on causing longer execution times of the entire system. The architecture requires a test run to abstract information, where the parameter range of the to be controlled actuator is altered. In order to determine which information is more relevant than others, the production system should explore its own range of actions. After this test run the system analyzes the recorded sensor data in order to find representative features. It may solve feature calculation separately for different kind of sensors, however the sensory units should ideally be trained to map the sensory input into the learned feature space. Finding a useful feature space representation is critical, because the system will only be able to recognize or react to changes of the features. The way of feature selection may vary for different production processes, in the last section, a best practice experimental experience for laser beam welding is demonstrated. Secondly, the step of supervised learning will be discussed. In natural human cognition, for instance in childhood, we often learn from our parents to manage complex tasks. A machine should have the possibility to learn its task initially from a human expert. In industrial production a qualified human supervisor is usually present when the production system is being installed or configured. The architecture of the present invention uses hu- man-machine communication to receive feedback from an expert, for instance, a graphical user interface. As mentioned above, in this architecture at least one test action per actuator or a test run is needed in an initial learning mode. During these tests, the robot executes one actuator from minimum to maximum output and the sensor data input is stored. After this run, an expert gives feedback as to whether the robot executed the actuator correctly or whether its action was unsuccessful or undesirable. The feedback may have many different categories, so kinds of failures and exit strategies may be defined. A classification technique may then collect the features with the corresponding supervisory feedback. Together with lookup tables the classifier will serve as knowledge and planning repository and for classification of the current system state. The Support Vector Machines, Fuzzy k-Nearest Neighbour, and Artificial Neural Networks will be discussed later as classification techniques. The more the human expert teaches the machine the likelier the system will achieve the desired goal. For cost saving the necessary human supervisor time should be minimized. Thus, if no feedback from an expert is available, the architecture may use after the first test run its learned knowledge to give itself feedback, as explained in the unsuper- vised learning paragraph. Thirdly, the step of adaptive process control should be discussed. To realize a robust process control in an industrial production process a fast closed-loop control is often required. If you think of a joining or cutting process, the loop should be completed at least once before the interaction zone has left the processing area. In a laser power controlled welding process for instance, a real-time process control implies that the laser focal spot should not have left the previous interaction zone, where the control-loop was started from. The advantage of the architecture of the present invention is that the use of features instead of raw sensor data permits to complete control-loops faster while the loss of information is minimized. In this architecture any kind of controller design may be implemented, which fits to the classification output. A simple version would be to have two possible classification output values: too much and too low. A PID controller could adjust a parameter of the system's actuators, as it will be the case in the last section.

Fourthly, the step of autonomous process control and unsupervised learning will be discussed. While a production system is executing an adaptive process control with the dis- cussed architecture, it may be the case that the system experiences something new in terms of not previously learned. This may be the case in assembly lines if work loads change or any other process parameter not recognized by the system is altered. A novelty check based on the trained data may detect such differences. In this architecture it would result in a change of system mode either to supervised learning if a human expert is present or to autonomous process control. Thus, the system may also try to solve the problem with a self-learning mechanism. In the remainder of this paper a mapping of characteristics is described as one possibility. Another possibility would be to integrate reinforcement learning.

As described above, the architecture abstracts information which is reduced in data vol- ume. What has been called 'activation patterns' in the section concerning the inspiration from natural cognition, could also be understood as features representing sensory events. Using, for instance, dimensionality reduction, a lower dimensional feature calculated from the training events would indicate if the system has experienced a certain event. A classification with Support Vector Machines can categorize and distinguish the events. After an initial feature calculation, the presence of features in sensory data inputs can be processed a lot faster than processing raw sensory data. A combination of data reduction and classification techniques is promising for process control because it enables it to act fast.

The architecture of the present invention proposed for cognitive technical systems enables several cognitive capabilities, such as obtaining relevant information, learning from a hu- man expert and reacting to new situations based on previously learned knowledge. This architecture may be used for different kinds of systems that need to control one or several actuators on the input of a high amount of sensor data. Compared to some other approaches, the learning and reacting capabilities seem to be limited; however, the architecture is very robust in terms of data acquisition; it is easy to use and can be realized for fast computing up to real-time closed-loop control of complex systems such as the application shown in the following section.

In the last section, an illustrative embodiment of the present invention being related to a technical cognition and adaptive laser welding process control is described.

A typical procedure for production systems is that a unit of an assembly line must first be configured and then be monitored for quality assurance. This is also the case for laser welding. When materials are processed with the use of laser light, a high degree of precision is necessary. On the other hand, welding with laser beams is hard to observe because of strong radiation and process emissions. For these reasons many different sensors are used for monitoring activities. Yet even then it is challenging for human experts to tell whether a welding action was successful and will result in stable welding seams by evaluating the monitoring results. In industrial production, these processes are usually initially configured with many manual trials, resulting in high costs in labour and machinery. All process parameters are kept constant, because change would result in high recalibration costs and may cause production to stop. A cognitive system for laser material processing capable of reacting appropriately to changes would be of great help and an economic upgrade.

An approach for a cognitive system for laser material processing following the architecture of the present invention is shown in figure 2. The data processing is structured within this architecture. The following sensors are used: a high-speed camera, sensors for solid-borne and air-borne acoustics, and three photodiodes recording different process emissions in different wavelengths, as demonstrated in figure 3A and 3B. As mentioned above, the cognitive capabilities under evaluation can be separated into four categories: first, abstracting relevant information; second, learning from a human expert how to control; third, acting on learned knowledge; and fourth, autonomous decision-making in previously unknown situations.

Firstly, the step of Abstracting Relevant Information should be discussed. For the cognitive laser material processing system, we introduced cameras, photodiodes, and sensors for solid-borne and air-borne sound, offering a lot of valuable process information. This sensory data has to be recorded in an initial test run altering the applied laser power. Other actuators could be altered as well, in this case laser power changes have negligible reaction times but great influence for the resulting weld seam. Dimensionality reduction mainly serves as feature extraction technique. Linear and non-linear dimensionality reduction en- ables us to map a high-dimensional observation space Y to a low-dimensional feature space X. All of the data from the applied online sensor system can be represented by a 26.603 dimensional vector, which can be reduced to only 33 dimensions, representing the relevant process-information by combining different dimensionality reduction techniques. Afterwards, it is possible to embed the sensor data input to the learned feature space and gain the current system process state by out-of-sample extension. Experiments have shown that every sensor ideally has an optimized combination of feature extraction.

Secondly, the step of Supervised Learning will be discussed. Laser material processing systems are usually set up and their processes configured by human experts. The discussed architecture may simplify and accelerate this process. When the system performs a test action such as a laser power ramp for welding a human expert points out, over a graphical user interface displaying the workpiece, how the processing result would be classified for the different workpiece areas. For instance, the expert may mark a poor weld with not enough laser power, a good weld, and a poor weld with too much laser power applied. Our system stores this information together with the extracted characteristics described above with a classification technique. Several classification techniques have been evaluated and have been found to be suitable embodiments of the present invention, such as Support Vector Machines, Artificial Neural Networks, and Fuzzy k-Nearest Neighbour. All of the named classifiers achieved good results; the extracted characteristics seem to be separable from each other for many different process setups. The combination of feature extraction and classification proofed to be a reliable monitoring tool for stainless steel welding, for example soiled areas on workpieces or variation in material thickness may be monitored. Even previously hard-to-detect faults can be monitored, such as a too large gaps in between two sheets to be joined, as shown in figures 4 and 5.

Thirdly, the step of Adaptive Process Control will be described. Different closed-loop con- trol experiments were performed on the basis of the classification probabilities. Combining PID-control with the classification results enabled performance of laser power controlled overlap welds of stainless steel workpieces, compensating for the focus shift in processing optics by varying the z-axis, and reducing the influence of the gap between the parts to be joined. Using lower data volume features and simple class structures permitted perform- ance of laser welding process control in real-time. According to the present invention, closed-loop control cycle frequencies greater than 500 Hz could be achieved, including feature mapping and classification of sensor data from camera pictures and photodiodes.

Fourthly, the step of Autonomous Process Control and Unsupervised Learning is discussed. Even though every attempt was made to keep all processing parameters constant for a configured process, influences may occur in varying workpieces, such as changes in mounting and workload. According to the present invention, a self-learning mechanism is integrated. A novelty check on the basis of the trained features can detect new or previously unknown situations. In this case, the system performs another test action and classifies the new workpiece with the previously trained features. This time it does not need to consult a human expert; it can map the gained knowledge onto the new workpiece autonomously and adjust the process control appropriately. This cognitive capability has been tested for varying material thickness of ±15%; the results are demonstrated in figure 6. The usual operation area is indicated within this figure, producing good and poor welds. Especially within this area, the trained feature sets are now identified earlier or later on the workpiece with different material thickness. The results show that the mapping of charac- teristics seems to work reliably, because the mapping points out the necessary laser power adjustments. The experimental results indicate that the system of the present invention is capable of making a judgment and adjusting the applied laser power for receiving the desired welds accordingly. Further techniques for this step could be employed, such as re- inforcement learning in order to increase high-level adaptivity.

In summary, an architecture of the present invention is presented, which is suitable to different process control task within industrial production. The design inspired by natural cognition is applied to laser welding process control problems. The method and the apparatus of the present invention provide cognitive capabilities for managing multiple sensors and actuators. The system has been able to learn from a human expert how to weld materials, make decisions in previously not trained situations, improve monitoring and process control with 500 Hz and greater.

In order to execute the present procedure it is necessary to store the dimensionality reduced sensor data from the initial machining or processing processes. These data entries are labe- led with the human expert feedback and therefore grouped within classes. When performing the process control the system classifies the incoming dimensionality reduced sensor data by calculating the class probability of the existing data entries or features. This step provides a similarity probability within the feature space. It therefore determines the process status and its probability of being in a desired state. The system optimizes this proba- bility of being in a desired state. The good class probability can be taken as set value for a controller such as PID, which tries to achieve 100%. Further human expert feedback can increase the labeled data stored within the classifier. Within this step the data of the current machining or processing process is labeled with the human expert feedback and is additionally stored within the classifier. This allows to support the machine if it cannot get into a good class system state or to fine-tune the performance.

A number of different methods such as Kernel Principal Component Analysis (PCA), or Artificial Neural Networks, or Support Vector Machines and similar classifiers allow the machine to do a novelty check. With the novelty check the system knows if the system state exists within the learned data entries. It may be realized with another classification. It can actually determine if the system's state is similar to the previously learned dimensionality reduced sensor data. For instance all labeled data entries within the feature space are labeled for another classification as known data entries. With a Support Vector Machine classification the known data entries may be one class and the feature space origin another class. The Support Vector Machine then calculates the probability if the system's state is known or similar to previously learned data entries. If this probability is below a preset threshold, the system's state may be considered as new or unknown. The novelty check provides safety, because the system executes the desired action of a human expert only on learned knowledge. In unknown states the system acts on similarities, which may not be appropriate for the desired machining process.

This novelty check can be used as another monitoring signal for instance. The human expert can trust all machining processes, where the novelty check has not been positive. In the case of a positive novelty check the system can either ask for additional human expert feedback or it can also start a self-learning process such as reinforcement learning or workpiece mapping. Other sensor data feedback may be used as reward function for a reinforcement learning agent. Workpiece mapping describes a process where dimensionality reduced sensor data is gained from a new machining process and labeled by feature similarity of previous features.

Despite other process control procedures the present invention allows to perform an adap- tive process control on a black-box model. It is not necessary to preset a process control model, where parameters can be adapted. The system can furthermore supervise itself using the novelty check and perform action upon it. It is especially suitable to processes with noisy and complex sensor data such as laser welding or laser cutting.

Claims

1. A method for processing workpieces by means of a cognitive processing head, comprising the steps of

performing an initial processing process of an initial workpiece type by acting on learned knowledge; and

making an autonomous decision in a previously unknown processing situation for adapting the learned knowledge to a secondary processing process of a secondary workpiece type.

2. The method according to claim 1 , further comprising the steps of

abstracting relevant sensor information concerning the initial processing process of the initial workpiece type; and

learning from a human expert how to control the initial processing process of the initial workpiece type.

3. The method according to claim 1 or 2, wherein processing of the workpieces comprises a laser welding or cutting process.

4. The method according to claim 2 or 3, wherein abstracting of relevant sensor information comprises an initial test run, in which an actuator parameter is altered and the sensor information is processed by a feature extraction technique to embed the sensor data input to a learned feature space.

5. The method according to claim 4, wherein the relevant sensor information comprises information of sensors including a high-speed camera, sensors for solid- borne and air-borne acoustics, and three photodiodes recording different process emissions in different wavelengths.

6. The method according to claim 4 or 5, wherein the feature extraction technique comprises a linear or non-linear dimensionality reduction processing of the sensor data input.

7. The method according to one of the claims 2 to 6, wherein the learning from a human expert how to control an initial processing process comprises controlling of actuators including a linear drive or robotic positioning actuator or an actuator for setting a laser power.

8. The method according to one of the claims 2 to 7, wherein learning from a human expert comprises classification techniques such as Support Vector Machines, Artificial Neural Networks, and Fuzzy k-Nearest Neighbour.

9. The method according to one of the preceding claims, wherein performing an initial processing process comprises a combination of a PID-control with classification results.

10. The method according to one of the preceding claims, wherein performing an initial machining process comprises performing laser power controlled overlap welds, in which too large gaps between two sheets to be joint are monitored.

1 1. The method according to one of the preceding claims, wherein making an autonomous decision includes a novelty check on the basis of the trained data.

12. The method according to claim 1 1 , wherein, if the novelty check is positive, a further test action is performed to classify the secondary workpiece type with the pre- viously trained features by means of a supervised or unsupervised learning procedure.

13. The method according to claim 12, wherein the secondary workpiece type is related to a different material thickness of the workpiece.

14. The method according to claim 12 or 13, wherein the supervised learning procedure includes reinforcement learning.

15. A cognitive processing head, comprising a control unit being adapted to perform a method as claimed in one of the preceding claims.